LIBRARY

SCHOOL
BEY, CALIFORNIA 93943-5003

)NTBB

OHTEJ

L POSTGRADUATE SCHOOL

Monterey, California

I titL

SPEECH RECOGNITION IN A

COMMAND AND CONTROL
WORKSTATION ENVIRONMENT

by

Michael A. LeFever

March 1987

Thesis Advisor

Gary K. Poock

Approved for public release; distribution is unlimited.

T231306

UNCLASSIFIED

SEC'jai

ClaSS. c CaTOn 0' T^S paG:

REPORT DOCUMENTATION PAGE

U REPORT SECURITY CLASSif'CATiON

UNCLASSIFIED

b RESTRICTIVE MARKINGS

la SECl>R'Tv Classification authority

:D OEClASS-? CAT. ON - DOWNGRADING SCHEDULE

3 D:STRiauTiON/ AVAiLAS'LiTY OF REPORT

APPROVED FOR PUBLIC RELEASE;
DISTRIBUTION IS UNLIMITED

PERFORMING ORGANIZATION REPORT \uM3ER(S)

5 MONiTOR. NG ORGANIZATION REPORT NUMBER'S)

ca NAME Cc PERFORM. NG ORGAN. ZAT ON

iaval Postaraduate School

to O^'.CE S ' M 8 O i
(!f applicable)

55

a NAME OF MON.TORAG ORGA.VZAT O

Naval Postgraduate School

Sc ADDRESS .City. State, and Z'.PCode)

Monterey, California 93943-5000

7b ADDRESS (Cry. Stare, and ZIP Code)

Monterey, California 93943-5000

IA.VE OF FuNOiNG SPONSOR:N<
IRGAN ZAT ON

So DFFiCc SYMBOl

f/f applicable)

:nt nst«u'

3c AODRESSfGry 5r<jre W Z7F> Ccxi*;

iO'JRCE OF funOin;

PROGRAM

Element no

PRO.-EC

NO

'AS<
NO

ACCESSC-

NO

M * Tl£ (include Security Clarification)

SPEECH RECOGNITION IN A COMMAND AND CONTROL WORKSTATION ENVIRONMENT

-ERSO'-V aut-or;s>
.EFEVER, MICHAEL A

•J 'V?t Z' *£PO«T

Master's Thesis

V. E C C '. E 3 E 3

► aov

?A*E OF =E.PQRT {Year Month Day,

March 1587

75

6 5LPpLEV£NTARY NO'ATiQN

cosa: codes

GROUP

SUB-GROUP

'8 SUBJECT TERMS .Continue on reverse if necessary and identify by O'cc* number,

SPEECH RECOGNITOR COMMAND AND CONTROL WORKSTATION,

CCWS, SRI "BERKELEY1 SPEECH BOARD, VOTAN SPEECH RECOGNITION

9 A8SrRACT 'Continue on reverse if necessary and identify by blo<k number)

This thesis investigates speech recognition in a command and control workstation
environment. It discusses the Navy's need for a command and control workstation
(CCWS) and the importance of the human interface desian. In particular, it evaluates
the performance of Stanford Research Institute International (SRI's) 1000 word
discrete speech recocnizer. The speech board is intended to be used in the Command
and Control Multi-Media workstation being developed by SRI. Additionally, it
investigates a VOTAN continuous recognizer (currently in use by research and
commercial vendors) in an interactive warfare simulation game. The results indicate
that
inpi
disi

20 0 S"R"3UT ON; AVAILABILITY OF ABSTRACT

5(3 -NClASSiF'EDHjNI'MiTED D SAME as rpt q otic LSERS

21 A8STRACT SECURITY CLASSIFICATION

UNCLASSIFIED

'.2d NAME OF RESPONSIBLE ND'ViDUAl

22b TELEPHONE (include Area Code) 22c OfUCE SvM3i

GARY P00CK

(408) 646-2636

55Pk

DDFORM 1473.84MAR

I3 APR eO't'On rray be used until e«f-austed
All other editions are obsolete

1

secl,Rity classification of th.s race
UNCLASSIFIED

Approved for public release; distribution is unlimited.

Speech Recognition in a

Command and Control

Workstation Environment

bv

Michael A. LeFever

Lieutenant Commander," United States Navy

B.S., United States Naval Academy, 1976

Submitted in partial fulfillment of the
requirements lor the degree of

MASTER OF SCIENCE IN SYSTEMS TECHNOLOGY
(Command, Control and Communications)

from the

NAVAL POSTGRADUATE SCHOOL

March 1987

ABSTRACT

This thesis investigates speech recognition in a command and control workstation
environment. It discusses the Navy's need for a command and control workstation
(CCWS) and the importance of the human interface design. In particular, it evaluates
the performance of Stanford Research Institute International (SRI's) 1000 word
discrete speech recognizer. The speech board is intended to be used in the Command
and Control Multi-Media workstation being developed by SRI. Additionally, it
investigates a VOTAN continuous recognizer (currently in use by research and
commercial vendors) in an interactive warfare simulation game. The results indi
that speech recognition systems could increase the capability of the commander to
input and access information, provide more rapid response to information desired or
displayed, and enhance human interaction in the man-machine interface. Past, current.
and future speech applications are discussed.

TABLE OF CONTENTS

I. SPEECH RECOGNITION IN A COMMAND AND CONTROL
WORKSTATION ENVIRONMENT 9

A. INTRODUCTION • 9

B. PURPOSE OF HIE THESIS 9

C. SUMMARY ' 10

II. ARCHITECTURAL REQUIREMENTS FOR A COMMAND

AND CONTROL WORKSTATION (CCWS) 11

A. OVERVIEW 11

B. THE COMMANDER S DECISION PROCESS 12

C. NAVY NEED 13

1 . Brief History 13

2. Current Deficiencies in U. S. Naw Information

Processing ' 14

3. Systems Approach 14

D. SOLUTIONS: A NATIONAL LEVEL DISTRIBUTED
COMMAND SUPPORT (DCS) SYSTEM 15

E. FLEET AND BATTLE GROUP LEVEL: CCWS 17

F. HUMAN INTERFACES TO CCWS 18

1. Voice Entry 20

2. Automatic Speech Recognition Requirements 20

G. CONCLUSION 20

III. SPEECH TECHNOLOGY PAST. PRESENT AND FUTURE 22

A. OVERVIEW7 22

B. DEFINITION OF TERMS 22

C. PAST 24

D. PRESENT 25

1. Overview 25

2. Speech Applications in Command and Control 26

E. FUTURE 26

IV. TEST. ANALYZE. AND EVALUATE THE SRI BERKELEY'
SPEECH BOARD 29

A. DESCRIPTION 29

B. THE SUN-170 MICROSYSTEMS WORKSTATION 29

C. MARA 30

D. THE SRI 'BERKELEY' BOARD 30

1. Front End 30

2. Comparator 31

E. . SUBJECTS 31

F. TRAINING ALGORITHM 31

G. THE VOCABULARY 32

H. PROCEDURE AND DATA COLLECTED 32

I. RESULTS 34

1. Accuracy 34

2. Interoperability of Voice Patterns for Different Users 34

3. Accuracy in a Noisy Environment 36

J. SYNTACTIC FEEDBACK 36

K. CONCLUSIONS AND RECOMMENDATIONS 37

V. TEST. ANALYZE. AND EVALUATE A COMMERCIAL
CONNECTED VOICE RECOGNITION SYSTEM IN A
WARGAMING ENVIRONMENT 38

A. DESCRIPTION OF THE NAVAL WARFARE
INTERACTIVE SIMULATION SYSTEM (NWISS) .39

B. SCENARIO 40

C. VOTAN SPEECH RECOGNITION SYSTEM MODEL

6050 SERIES II 41

1 . Vocabulary Size 41

2. Programming 42

3. Operation 43

4. m Training Algorithms 44

D. SUBJECTS 45

E. THE VOCABULARY 45

F. PROCEDURE 46

G. RESULTS 49

H. CONCLUSIONS AND RECOMMENDATIONS 51

VI. CONCLUSIONS 52

APPENDIX A: SRI 100 WORD VOCABULARY 54

APPENDIX B: 240 WORD VOCABULARY 55

APPENDIX C: SCENARIO BRIEFING 59

APPENDIX D: VOTAN VOCABULARY FOR NWISS 64

LIST OF REFERENCES 70

BIBLIOGRAPHY 72

INITIAL DISTRIBUTION LIST 73

LIST OF TABLES

1. ' SRI 1000 WORD RECOGNITION PERFORMANCE 34

2. SRI TI DATA BASE PERFORMANCE (ERRORS OF 320) 35

3. NFS 100 WORD VOCABULARY TEST 35

4. NTS 240 WORD VOCABULARY TEST 35

5. INTEROPERABILITY TESTS 36

6. NOISY ENVIRONMENT : 36

7-. INITIAL TEMPLATE LISTING 4^

S. REVISED TEMPLATE LISTING 4S

9. SRI VS VOTAN 240 WORD RECOGNITION ACCURACY TEST 49

LIST OF FIGURES

1.1 Distributed Command Support 16

1.2 Functional View of the Commander's Decision Process Supported bv

ccws : is

1.3 CCWS Man-Machine Interface Components (Adapted from Poggio,

1985) 19

2.1 Automatic Speech Recognition Systems 23

5. 1 NTDS Symbology 40

5.2 Configuration To Run NWISS With The VOTAN 41

I. SPEECH RECOGNITION IN A COMMAND AND CONTROL
WORKSTATION ENVIRONMENT

A. INTRODUCTION

Ever since the rapid influx of microcomputers, there has been increasing incentive
to enhance the productivity of humans. The job of automating routine tasks, acquiring
and communicating information, and the very popular intelligent support of decision
making are all attempting to exploit the potential of these machines. Of equal
importance is the growing effort to enhance the productivity of humans through man-
machine interfaces to take advantage of these growing capabilities.

We can exchange information in a variety of methods. Our most efficient
communication should be available when we want to communicate information via a
computer. It has long been known that speech is the most natural and fastest form of
communication for us and therefore, should be considered as the unrivaled interface for
system optimization.

Research into automatic speech recognition systems has been ongoing for over
thirty years. Automatic Speech Recognition (ASR) is defined as the ability for the
computer or device to correctly recognize spoken words and translate that into a
predetermined output string to the computer. There are many advantages of using
voice input. The most important of these characteristics are freeing the user's hands
and eyes for other tasks, employment in low light or dark areas, and the freedom of
movement from a specified location.

From this list of advantages, it would be easy for us to let our imaginations
wander and generate a listing of thirty or more applications for voice input. Quality
control on assembly lines, sorting of packages, office automation, aircraft control,
disabled control of wheel chairs, and many more well suited examples could be
enumerated. The focus of this work is to examine speech applications in the area of
Command and Control and in particular a Command and Control Workstation
(CCWS).

B. PURPOSE OF THE THESIS

Even though there have been over 30 different theses accomplished at the Naval
Postgraduate School related to speech recognition systems alone, there is little

awareness of speech applications in the naval environment. Evaluating state of the art
systems and recommending various areas for speech applications in a shipboard
environment may raise the awareness of this technology and help to incorporate speech
technology in the future designs of man-machine interface. It is without a doubt an
area of technology that has far reaching consequences for the commander in the
growing age of computers.

C. SUMMARY

This thesis describes the purpose of the CCWS in the Distributed Command
Support System and the key role of speech in the human interface. Basic speech
technology past, present, and future is described in Chapter III. A description of the
experiment used to test the Stanford Research Institute International (SRI) 'Berkeley'
1000-word discrete speech recognizer is presented in Chapter IV. A follow-on
experiment utilizing a commercially available VOTAN continuous speech recognizer is
described in Chapter V. Finally, conclusions from these experiments and the author's
recommendations for additional speech applications in a Command and Control
environment are offered.

10

II. ARCHITECTURAL REQUIREMENTS FOR A COMMAND AND
CONTROL WORKSTATION (CCVVS)

A. OVERVIEW

This chapter will investigate the needs and the architectural requirements for a
Command and Control Workstation (CCWS). The particular workstation this paper
will investigate is the SUN Microsystems Computer Model- 170 proposed by Stanford
Research Institute (SRI) for the U. S. Navy needs. This paper will develop the
architectural framework needed above the workstation system and focus on the
requirement to include well engineered human interfaces. This is motivated by the
immense amount of information flow that this future system will support. Voice
recognition is examined as a potential solution to the growing complexity of getting
information to the commander.

In every C^ system there is a commander who sole purpose is to make timely and
knowledgeable decisions. An understanding of the commander's decision process is
essential to ascertaining what the CCWS must support. Ever}' new technological
advance alters the balance of forces and must be carefully considered. The CCWS
design seeks to create an advantage by integrating a multitude of sources into one
system. The commander must be able to exercise control over these combined
resources. He must obtain the various data in a form that he can best digest. This is
not a trivial problem as the amount of information available to him can quickly
overwhelm his staff and work against their objective. It requires a systems approach in
solving the problem of fusing these composite sources o[ data. In any systems
approach one must understand how the system will compliment the architectural level
above and the layer below. We will begin with the definition of some relevant terms
and an examination of the processes and structures germane to the CCWS.

Much has been written to define Command and Control (C ) in various sources.
In lieu of adding another definition to the growing mass we will use the Joint Chief of
Staffs Dictionary to delineate C .

The exercise of authority and direction by a properlv designated commander over
assigned forces in the accomplishment of his mission. Command and control
functions are performed through an arrangement of personnel, equipment,
communication, facilities, procedures which are emploved bv a commander in
planning, directing, and coordinating, and controlling forces" and operations in
the accomplishment of a mission. (JCS, 19S4)

11

As defined in Dupuy (1986), command is the authority vested in an individual of
the armed forces for the direction, coordination, control, and administration of military
forces, and control is the authority exercised by a commander over the activities of
subordinate organizations or entities. Since a computer is the heart of CCWS, we will
interpret computer as a machine which performs electronic, mathematical manipulation
of new inputs and existing data to obtain useful outputs in near-real-time.

In the simplest terms Command and Control is a process by which a commander
directs his resources to achieve a goal. One of the key resources is the information
increasingly provided by a system of computers.

B. THE COMMANDER'S DECISION PROCESS

The commander's primary7 goal is the accomplishment of the mission. He must be
able to assimilate copious amounts of information and data. Based on his
understanding of tiie situation he must then make the split-second decision for which
he alone is accountable. The process can be thought of as a continuous loop which
observes the effect of the decision on the environment. This outcome will be reflected
in the data or information obtained, and the process repeats.

There are many models depicting this reiterative decision process or loop: J.
Lawson's model, Boyd's OODA loop mentioned in Orr (1983) and the SHOR
paradigm mentioned in Wohl (1981). All of these illustrations are merely extensions of
the stimulus response model of classical behaviorists. For simplicity and to align this
feedback loop to the basic functions of a shipboard Combat Direction Center, our view
of the maritime commander's decision process will be:

• COLLECT--to obtain combat information from all available sources

• PROCESS--to sort, review, appraise, and correlate all information

• DISPLAY— to present the information that best serves the decision maker

• EVALUATE-dccide

• DISSEMINATE-distribute the decision

This decision process can be at any level in the command structure. For the
Commanding Officer of a ship, Battle Force Commander, or even the Fleet
Commander, the process is the same. These loops are nested within each other
forming a hierarchy. The systems and processes that makeup these nested loops all
work toward supporting the commander in directing and controlling his forces. The
design of a C" system must support all these processes in a timely and accurate

12

manner. To motivate an understanding of the factors needed in today's information
systems, we will briefly examine the background leading to the current dilemma in
information management for the U.S. Navy maritime commander.

C. NAVY NEED
1. Brief History

A primary input to the commander's decision process is monitoring the
environment. This process within the decision loop is supported by proper
management of his sensors (collection), the processing of this information (process),
and presenting the information useful to the decision maker (display). Historically, the
tactical commander relied solely on the organic sensors of his battle group.
Information from the Fleet Command or other sources was spotty at best. In the 40's
and 5u's', the technological improvements in sensors and communications equipment
produced a huge amount of information for the commander. There was an early
indication that the unsupported decision maker could easily be overwhelmed. As
pointed out by G. A. Miller (1956) in a psychological review ". . . current manual
methods oC information processing incident to decision making may be inadequate, and
new types of filtering and preplanning will be required." (Wohl, 1981)

Through the years following, the need for a device to assist the commander
became even more apparent. The technological advances in computers, automatic data
processing and weaponry were overwhelming. The effect of longer range and faster
aircraft, missiles, and guns was to greatly increase the area of responsibility for the
commander. The protection of his force utilizing the ' Defense-in-Depth' concept,
consisting of a surveillance area, engagement area, and a vital area, was degraded by
his inability to manually track all the contacts in these areas effectively. Our systems
were quite inadequate to fully support the decision maker. Even the Soviets realized
this dilemma as evidenced in this quote from the General of the Army S.M. Shtmenko.
U.S.S.R :

The volume of information that staffs must process has increased many fold since
World War II and the time allowed for decision making has decreased many fold.
As a result the requirements on the "brain capacity" "of commanders and" stalls
have increased vastly. To meet these requirements by simplv expanding, the
administrative apparatus is fundamentals impossible . . .' . The "only escape from
this incompatible situation lies in the extensive application p( automation,
pnmarilv computers ... a "man-machine" system is more perfect than "man"
alone of "machine" alone .... Information technoloav does not simplv help the
commander and his staff, but also stimulates the '"development of 'collective
militarv creativitv. in which the largest group of people, including those separated
by great distances, can participate. w(Druzhinin, 1972)

13

2. Current Deficiencies in U. S. Navy Information Processing

The U.S. Navy realized by the end of World War II that the current combat
information center (CIC) was quickly becoming outdated. The early 1960s saw the
first digital computer, Naval Tactical Data System (NTDS) operational in the fleet.
This was a revolutionary step. A machine had been connected, through
communications links, to another machine to pass real-time information in a
operational environment. NTDS is an automated method of collecting, processing,
displaying, and disseminating tactical information. Information is displayed
graphically, in real-time and provides the shipboard decision maker with a considerable
amount of information to direct his weapon employment. As time progressed, there
were tremendous advantages realized in obtaining information from other than organic
battle group sources (e.g.. national level sensors). This led to many ad hoc
improvements to the system that were outside of the original architecture for N'l DS
and were never really designed to interface with the system. Naturally, saturation
became a problem. There were so many diiTerent systems that often sailors were
required to accommodate the differences in data format, information fusion, and
sanitization of highly classified data and sources. The problem was summarized in
Local Command Center Network Statement of Work (LCCN) (197S) as follows:

The introduction of each new technology development (communications,
weapons, sensors, electronic warfare), whether bv enemv or (riendlv forces, mav
significantly alter the manner in which multiple platforms (ships,' aircraft and
submarines') can be most effectively coordinated. The proper exercise of
command and control in this changing environment requires that the combined
sources of data be presented to the "commander in a form which is tailored to his
resources, mission, and surrounding environment.

3. Systems Approach

The ad hoc solutions to these problems of coordination and interfacing were
complicating rather than supporting the commander's decisions. The increased
sophistication of existing systems and the addition of new requirements have caused
the individual number of components in systems to drastically increase. A systems
approach to effectively manage and assess the expanding individual systems becomes
quite evident.

The large quantity of information from national, joint, and. or Navy sensors is
indispensable to the commanders in the field. The extended battle group surveillance
area has grown proportionally to the range of the over-the-horizon weapons (both the

14

enemy's and our own) and global sensors. This vital data is available from many
sources, but the current flow of information makes it unobtainable. The information
that is available often requires manual correlation. As a result, the decision process
discussed above either lacks the necessary information or is overwhelmed by the reams
of unprocessed data. The intent of the Distributed Command Support (DCS) System
is to reduce the information processing and collection load through correlation,
tracking, and fusion of data.

D. SOLUTIONS: A NATIONAL LEVEL DISTRIBUTED COMMAND
SUPPORT (DCS) SYSTEM

As defined by Tanenbaum (1981), a distributed system is a special case of a
computer network with a high degree of connectivity, cohesiveness. and transparency.
It could be a stand alone system or on^ in which the data and information are
available to anyone in the network wherever they may be located. Its application in a
C~ environment has far reaching consequences.

The Navy understood its deficiencies in information exchange and the potential
in computer networks. The need for such a system was expressed in the following
Naval Need statement by Naval Ocean Systems Command (NOSC. 1985):

Existina and planned Navy Svstems (e.2.. sensors, communications, weapons and
C2 support svstems) are 'developed as"' stand-alone svstems. Coordination and
interpretation' between svstems is accomplished is a'n ad hoc. svstem unique
manner that often requires manual coordination. Advances 'in weapons.
surveillance and detection svstems are sisnificantlv increasina demands on the
Naw C2 svstems. Therefore, these svstems must be integrated in a more
adap'table, interoperable and survivable way.

The Distributed Command Support Svstem (DCS) will provide the command
centers with a more complete and overall combat picture from both afloat and
ashore sources. Throii2h DCS. commanders will be provided with the capabilitv
to extract information from data transfer svstems. combine that data with
artificial intellicence decision aids, and selectively present combat planning
decision aids usiiis communication protocols ....

The essence of the problem is the integration of a wide assortment of computers
and software. A non-degraded operation between systems as well as a stand alone
capability was envisioned. The DCS network as detailed by NOSC is shown in Figure
1.1

The DCS system is the integration of a wide variety of systems from an
assortment of users all able to share each others contributions to the network. Many
of these systems already exist with several planned for the near future (FY S7, SS).

15

The Local Area Network (LAN) is the heart of the system and has many different
methods to establish connections (e.g., satellite, high frequency, ultra high frequency
and Department of Defense Network (DDN)).

EXISTING/
PLANNED
SYSTEMS

Distributed Command Support

LINK-11 LINK-16 TACINTEL TADIXB

C2P

TADIXA

NTDS/
ACDS

POST

FDDS

NIL)

FLTBDCST CUDIX

MULTILEVEL
SECURITY

INTERFACE

NAVMACS

NIU

NIU

NIU

LOCAL AREA NETWORK

IBGTT

HP9020

CCWS

SYMBOLICS

WORK
STATIONS

ACDS Advanced Combat Data Systems
ANDVT Advanced Narrowband
Digital Voice Terminal
C2P Command and Control Processor
CUDIX Common User Digital Information

Exchange System
DAMA Demand Assignment Multiple Access
DDN Defense Data Network
FDDS Flag Data Display Systeml
FLTBDCST Fleet Broadcast
HFAJ High Frequency Anti-Jam
IBGTT Interactive Battle Group Tactical Trainer
LINK 1 1 Two-way Tactical Data Link NAVY

DDN

r

ANDVT

I

HFAJ

GATEWAY

DAMA MILSTAR
I
UHF

MILSTAR Military Satellite Communications System
LINK 16 Two-way Tactical DataLink AF
NAVMACS Navy Message Automated

Communications System
NIU Network Interface Unit
NTDS Navy Tactical Data System
POST Prototype Ocean Surveillance Terminal
TACINTEL Tactical Intelligence
TADIXA Tactical Data Information Exchange A
TADIXB Tactical Data Information Exchange B
UHF Ultra-High Frequency

Figure 1.1 Distributed Command Support.

Computer to computer systems cannot communicate unless they are compatible,
for instance, operating with the same protocols. If they are not, a scheme must he
developed to connect them and at the same time minimize the effect o( the changing

16

protocols on processing speed. "The key to DCS is the development of standard
application protocols that will support intra- and inter-platform computer to computer
tasking." (NOSC 1985) As shown in Figure 1.1, the NIUs or Network Interface Units
are used to convert the protocols of one system to be compatible with another. NIL' is
analogous to the gateway shown at the bottom of Figure 1.1. The difference is that a
gateway may be capable of connecting two or more networks.

E. FLEET AND BATTLE GROUP LEVEL: CCWS

As shown in Figure 1.1. the CCWS is an integral part of this network. This is
where the commander interfaces to the system and as such is the focus o[ the rest of
this thesis. It will receive all the information on the network. A secure computing
project will make it possible for all the users to have the same data base but have
access only to those data elements for which they have the security and need
requirements. The use of a trusted guard will control the access to the data base and
allow secure operation of the system with various levels of classification. For example,
the Fleet Commander may have global access and unlimited security eligibility while
the squadron commander will have theater coverage and security access for only
specific areas. The major advantage is that the entire data base will be in every
location increasing the connectivity and cohesiveness of the information.

The desirability of personal computing techniques utilizing a distributed
workstation environment for the support of command and control operations for the
U.S. Navy was formally initiated in early July 1980. SRI International was tasked with
a feasibility study. Computer systems and technology has significantly changed since
the initial study; however, the basic capabilities and design considerations have
remained intact. The capabilities of a workstation in a Command and Control
distributed network as pointed out by Poggio (1985) should be :

• The expeditious acquisition of up-to-date multi-media information

• Flexible, reliable, timely exchanse of information among people, and between
people and processes.

• Rapid match of information transport requirements to dvnamic communication
capacity

• Survivability - loosely coupled autonomous systems

These capabilities translate directly into the Distributed Command System and a
battle group environment. The intelligence gathered from outside sources would be
combined with the sensor information provided from the battle group's organic

17

equipments. The capability for many users to simultaneously plan, decide, and
disseminate information in a multi-media environment will greatly enhance the
commander's decision process.

In addition to network information, the system is designed to provide decision
support systems to aid the commander in the decision process. Refering to our model
of the decision process shown in Figure 1.2, one can readily see that the CCWS
supports all four of the five functions and assists the actual commander's decision.
Today's Battle Group Commander should have at his disposal all the available
information utilizing the technological hardware and software to make the correct
decisions or evaluations. Therefore, to support the commander we should allow the
computer to do what it can do best (i.e., fusion of data) and allow the human to do
what only he can do, make the decisions. SKI is incorporating these ideas into a
computer based multi-media information system. The current design of the
workstation plans to accommodate this arrangement.

— 1

COLLECTION

i

PROCESS

DISPLAY
EVALUATE

4
4

<

CCWS

SUPPORTS

ALL

THESE
ACTIVITIES

A DIRECTLY

<j INDIRECTLY (ASSIST)

DISSEMINATE

4

f

L_J

figure 1.2 Functional View of the Commander's Decision Process

Supported by ( CWS.

F. HUMAN INTERFACES TO CCWS

Everything meaningful in the operation, extraction, and manipulation ol
information available from CCWS results from human interaction with the display.

Since the sole reason for the workstation is to assist and extend the capabilities of the
commander, the user interlace should be of utmost importance. As stated in NOSC,
(19S5) ". . . the man-machine interface must be more natural and efficient, readily
adaptable to the peculiarities of the user and support multi-media (i.e.. voice, graphics,
text) messages and information." High resolution, bit-mapped color displays,
sophisticated window and cursor controls, and speech recognition are all available now
for implementation in these personal workstations.

Figure 1.3 taken from Poggio (19S5) shows SRI's design considerations for
several man-machine interface components. The various instruments by which we
communicate instinctively (speaking, pointing, and writing) are all available in these
human interfaces. The evaluation and enhancement of the man-machine interface,
particularly in the speech realm, is the focus of this thesis.

MONOCHROME
DISPLAY

( ]

•"» I

COLOR DISPLAY

SPEECH DIGITIZER'
SYNTHESIZER

KEYBOARD

Ficure 1.3 CCYVS Man-Machine Interface Components
(Adapted from Poggio, 1985).

19

1. Voice Entry

Since humans have such a propensity for talking, it is only logical that speech
input,' output would be one of the ideal man-machine interfaces. Automatic Speech
Recognition (ASR) is defined as the ability for the computer or device to correctly
recognize spoken output and translate it into a predetermined output string to the
computer. There are many advantages of using voice input. The most important of
these advantages is freeing the user's hands and eyes for other tasks, allowing for
increased productivity and more rapid system response because speech input is faster
than conventional keyboard entry. The incorporation of ASR enables the O system
to be a true extension of^ the commander's decision making ability utilizing current
technology, his organization, and its procedures.

2. Automatic Speech Recognition Requirements

The following is a list oi" the critical requirements necessary- in an automatic

speech recognizer for incorporation into the CCWS.

Large vocabulary (capacity > 1000 words).

Real-time response.

Very high recognition accuracy ( > 98%).

Adaptable to the user, (i.e., the user should not have to modify or alter his
speaking rate significantly)

• No deterioration in accuracy in noisy and stressful environments.

These specifications are believed by the author to be those items necessary for an

effective and viable speech recognition system. The minimum capacity of 1000 words

was specified since this was a previous goal set in 1971 by the Department of Defense.

(Barr and Feigenbaum, 1981) An accurate, versatile, and fast large vocabulary system

which adapts readily to any user should be the goal oi" all manufacturers of automatic

speech recognizers. Consequently, this list will be the criteria for final evaluation of

the SRI 1000 word discrete recognizer and the VOTAN continuous word recognizer.

Since each speech recognizer is different, it is crucial that those responsible for the

man-machine interface spend sufficient resources in defining the requirements of a

particular system and finding the correct speech system to match.

G. CONCLUSION

The sole purpose of a command and control system is to support the
commander's decision process. The current system (NTDS) is overwhelmed by the
amount of information it must process and is proliferated with ad hoc equipments that

20

were never really designed to be interfaced with this system. An inadequate system
exists for today's commander.

A systems approach utilizing the technological advances in distributed networks
and personal computing led to the development of DCS and CCWS. The workstation
in development will incorporate the latest in protocols and will focus on supporting the
operational commander. The system design is to take full advantage of the man-
machine interfaces. Since our fastest and most efficient means of communication is
speech, it is only justifiable, that the design of the CCWS should consider speech
input output interfaces. This will ensure that the architecture for the command and
control workstation is designed to be a true extension of the commander.

21

III. SPEECH TECHNOLOGY PAST, PRESENT AND FUTURE

A. OVERVIEW

This chapter will describe the basic types of speech recognition systems and a few
of the fundamental terms associated with these systems. The history of speech
input output systems and forecasts of the future of speech technology are discussed in
broad detail. It is important to realize that each automatic speech recognizer uses
different algorithms. The user must be thoroughly familiar with the particular system
to ensure that it is the correct equipment for the task and that proper training and
programming of the system has been achieved. A basic familiarity with the terms and
the types of speech recognition systems is essential in comprehending this rapidly
growing technological field.

B. DEFLNITION OF TERMS

Before discussing speech recognition systems, we need to define and discuss the
various generic types of speech systems. As shown in Figure 2.1 there are two major
types: speaker dependent and speaker independent. A speaker dependent systems relies
entirely on the user training the speech recognition system. The user speaks an
utterance (one or more words in a phrase) usually 1-5 times for each word or a
particular output string. The equipment translates the frequency vs. time output into a
normalized, digital matrix. Depending on the manufacturer, these may be manipulated
by some averaging algorithm or just stored as separate templates in memory or in a
data base. A template is the digital representation or matrix of the utterance which is
used by the device to compare against your spoken word. Each system uses different
algorithms to calculate the template and a thorough understanding of the algorithm
used by the device is required to maximize recognition through proper training.

When a particular utterance is spoken, it is compared against the template in
memory and if it is within a pre-established limit or threshold, the device performs the
function the user has installed on the system. If it does not meet the threshold level,
the utterance is rejected and nothing is sent by the recognizer. Additionally, there are
two other events which can occur: an insertion or a substitution error. An insertion
occurs when a recognition takes place due to spurious noise or an utterance other than
those that are legitimate entries in the data base. For example, if you said 'defcon' or

22

SPEECH
SYSTEMS

SPEAKER
DEPENDENT

SPEAKER

INDEPENDENT

DISCRETE

CONTINUOUS

DISCRETE

CONTINUOUS

CONNECTED

CONNECTED

Figure 2.1 Automatic Speech Recognition Systems.

a similar word NOT in your database and the system recognizes and outputs the string
for 'defense'. A substitution on the other hand occurs when your input utterance is
calculated as a closer match to a different template in storage, thus incorrectly
recognizing another word. Tor example, if 'defcon' and 'defense' ARE currently in the
database and the utterance 'defcon' produces the string 'defense'. (Pallett, 1985)

The speaker-dependent, template matching systems are the most common
systems on the market. A system trained to a particular individual can achieve
recognition accuracies of 90-99 percentile. On the other hand, a speaker independent
system contains algorithms which are robust enough for any individual to be correctly
recognized. Such a device requires no training since each word is represented by
templates which are an average of a wide range of different utterances selected by the
manufacturer. Depending on the size and limitations of the vocabulary, recognition
accuracies are slightly less than those experienced by the speaker dependent systems.

23

The goal of most speech recognition manufacturers and researchers is to develop a
large vocabulary recognizer which is independent of the user. (Poock, 1986b)

Each of these two categories is further subdivided into three separate categories:
discrete, connected, and continuous. A discrete system or isolated word system as its
name implies is one in which the user must pause for a predetermined time (about .1
sec) between consecutive utterances. The device establishes the start and endpoint of
the word. These utterances are compared to what is in memory and the output string
is sent once the recognizer has calculated the best match.

The connected speech system requires no pauses between utterances. The system
is continually checking what is spoken and what is in memory. As the word or phrase
is recognized, the device is loading that particular string into the output buffer. Once
the user pauses, the system unloads all that it has accumulated in the butter.

In contrast, the continuous system outputs the prescribed string immediately upon
recognition and does not wait for a pause from the user. Even though there appear to
be no apparent word boundaries, the device is able to calculate matches and produce
the output strings. This is much harder than discrete recognition since there are major
changes which occur in the pronunciation of words at the word boundaries known as
co articulation. These are differences in speech patterns not found in isolated or discrete
word pronunciation.

Manufacturers today are still not in agreement over exactly what constitutes the
difference between these last two types. As stated earlier each and every system is
different and must be thoroughly tested and analyzed to ascertain exactly what the
manufacturer is trying to represent in his literature.

C. PAST

Many of the larger technical companies like IBM. Philco-Ford, RCA, and Bell
Telephone Laboratories started research back in the early 50's and 60's. It was not
until the early 70's that the first products commercially available were offered by
Threshold Technology, Inc. and Scope Electronics. (Poock. 19S6b)

Concurrently in the early 1970's, the U. S. Department of Defense Advanced
Research Projects Agency (ARPA) funded a five-year program in speech understanding
research (SUR).

ARPA funded five speech projects and several subcontracts for developing parts
of speech-systems. Some q{ the major ARPA contractors produced multiple
systems during the five-year period: Work at Bolt. Beranek and Newman, Inc.

24

(BBN) produced first SPEECHLIS and then HWIM (Hear What I Mean),
building on earlier BBN research on understanding natural language. Carnegie-
Mellon' Universitv (C.M.U.) produced the HEAKSEY-I and URAVOX svstems
in the earlv development phase (1971-1973) and the HARPY and HEARSEY-II
programs bv 1976. SRI International also developed a speech understanding
program, pa'rtlv in collaboration with Svstems Development Corporation (SDCk
(Barr and Feigenbaum, 19S1)

The ARPA projects were all built for the purpose of developing a speech
understanding device, but they varied considerably in levels of difficulty, number of
speakers, ambient noise, etc. As a result of this effort there was considerable progress
made toward practical speech-understanding systems. One of the most important ideas
to surface from these projects was the influence of Artificial Intelligence {AT) research
and system architecture. The researchers found phonetic recognition was the most
promising answer to continuous speech understanding, but at the time they did not
have the computing power necessary nor was it as straight forward as initially
anticipated. Since the early success of speech recognition used template matching,
industry abandoned the harder track of speech phonetics.

D. PRESENT
1. Overview

Currently there are literally thousands of organizations in the United States
and around the world exploiting speech systems. From controlling robot arms on the
space shuttle to incorporation into children's toys, speech input output systems are in
daily use and are growing rapidly. Despite ARPA's efforts, up until now all the speech
systems have consisted of relatively small quantity vocabulary7 pattern matching or
template matching techniques. The better systems can be expected to have recognition
accuracies of better than 97%.

There are several periodicals like the Journal of The American Voice I 0
Society and Speech Technology Man' Machine Voice Communications which reflect the
latest in research, applications of speech processing, and product reviews. In fact in a
recent edition, there were 193 different companies listed providing various products
and or services in the speech field. Speech recognition today is extremely capable and
reliable and could be applied to thousands of areas with more awareness and
understanding of its benefits to both user and management.

25

2. Speech Applications in Command and Control

Application of speech recognition systems in a shipboard environment need
not stop with the CCWS. There are many other areas where using this technology
could be beneficial. In the Combat Direction Center, manipulating NTDS displays and
functions on these consoles by voice in conjunction with the trackball tab. computer
controlled action entry panel (CCAEP), digital data entry unit (DDEU), and category
select panel would allow users to more quickly disseminate information and result in
less operator fatigue. Data retrieval by the Commanding Officer or Tactical Action
Officer to display decision aids or threat matrices by voice could promote better
weapon or countermeasures selections. The automatic speech recognizer could allow
the commander to focus totally on the display.

Combat Direction Center is not the only area on the ship that could benefit
from speech recognition systems. A voice activated expert system for controlling
engineering propulsion plant casualties would greatly enhance the reduced manning
policy on the automated gas turbine powered ship classes. Remote activation of
damage control (DC) or firefighting equipment by personnel outside the damaged
space could reduce the risk of damage to sailors and equipment.

The list could continue. Salfer (1985) presents a more detailed analysis of
applications of ASR systems onboard the FFG-7 class ships which could be expanded
to include other classes of ships as well. The underlying reason for pointing out
various other areas for speech applications is to stimulate awareness and generate other
ideas for applications for this technology.

It is important to note regardless of how much faster or better a system can
work employing automatic speech recognition technology, if the user and management
do not have the motivation to examine such a system, this equipment like others would
have no hope for success.

E. FUTURE

Speech recognition in no way should be considered stagnant. Manufacturers and
corporations are more than ever wanting to reap the benefits of this technological field.
As the awareness and knowledge of this technology becomes more widespread
especially in man-machine interface, a greater proliferation of systems will be seen.

The new horizon for speech recognition systems is to move away from template
matching schemes to the more flexible phonetic recognition. The basis of phonetic

26

systems is phonemes the basic units of all speech. Once the system is trained on words
utilizing all the combinations of phonemes, the formulation of any word is possible.
For example this phrase, taken from Speech Systems Incorporated advertising literature.

continuous speech development toolkit

would look like this phonetically.

kantinyuasspichdivelapmentulk.it.

The phonemes are then converted by different syntactic and dictionary builders in a
computer which produce the correctly formulated string. At the 1986 American Voice
Input; Output Society (AVIOS) convention, there was only one vendor Speech Systems
Incorporated who was marketing a phonetic recognition system. It is the first
commercial system oi" its type. It is surely the trend of future speech
recognition understanding systems and it is one focus of the Department of Defense
funding.

In addition to industrial and university research, Defense Advanced Research
Projects Agency (DARPA. formerly ARPA). is sponsoring another multi-million dollar
contract titled Strategic Computing Program. A major part of the Strategic Computing
Program is the integration, transition, and performance evaluation of speech
technology. "The speech recognition portion of the Strategic Computing Program is
divided into two major areas: continuous speech recognition and robust, connected-
word recognition . . ." (Strategic Computing, 1985).

The aim of this program is to make continuous speech recognition a realization.
The major thrust would be in the area of phonetic recognition to deal with speaker
variation, large vocabularies, natural grammars, and real time response. In the area of
robust speech recognition, the objectives are to improve upon current system's capacity
to deal with variations and distortions of the input speech signal in severe acoustic
noise and physiological psychological stress found in military applications. (Strategic
Computing Program, 1985)

Increased use of computers in problem solving will demand more emphasis on
man-machine interfaces. Speech recognition will be that interface which makes the
computer a true extension of man. We communicate with each other by speech, so it
should only be expected we can do the same via a computer. This cursory look at
speech types and speech related terminology is meant only to familiarize the reader

27

with terms to be used later and to introduce the ever broadening future of speech
input/ output systems.

28

IV. TEST, ANALYZE, AND EVALUATE THE SRI BERKELEY' SPEECH

BOARD

This chapter describes a series of tests whose purpose was to confirm the voice
recognition performance of the SRI 'Berkeley' board as reported in Murveit (1986).
The results of the SRI study suggest that a 1000-word discrete speech recognition
system does not sacrifice accuracy despite the high processing speeds necessary for
large vocabulary recognition. Their report indicates that tiie Berkeley speech board
system achieved a recognition accuracy of over 90 percent for a 1000 word vocabulary
and over 99 percent for a sixteen word vocabulary. In addition this chapter will
examine the algorithms used by the speech board for initial template creation, voice
recognition, and error correction.

A. DESCRIPTION

SRI selected the 'Berkeley' board because it was the state of the art in large
vocabulary speech recognition. A recognizer of this type was a necessary requirement
in a CCWS for a faster and more natural man-machine interface in command entry
and database access. Specifically, the research conducted by SRI was for the
enhancement of speech interfaces for natural-language data-base-management tools.
In cooperation with U.C. Berkeley. SRI modified the design slightly and interfaced it to
the SUN- 170 Microsystems computer.

B. THE SUN- 170 MICROSYSTEMS WORKSTATION

The SUN- 170 Microsystem workstation is a UNIX based computer system.
These workstations are used in a variety of applications. The value of workstations
was realized with the increase in computer power provided by the development of 16
and 32 bit microprocessors. A typical workstation will generally consist of a 1 MIPS
(million instructions per second) CPU, 2-4 Megabytes of memory, a high resolution
(1000 by 1000 pixels) display, a keyboard, and a mouse. The speech board is interfaced
to the SUN and receives the audio input directly.

The workstation used in this experiment is the host computer on the Department
of Defense Network (DDN) at address SRI-BOZO. There are several inherent
attributes like file transfer protocol (FTP) and telenet (TN) resident on the DDN
network which allowed remote work on the vocabulary and data processing from the
Naval Postgraduate School.

29

C. MARA

MARA is the hardware and software components that integrate the speech
recognizer into the workstation. The MARA system consists of:

• the computer and its programs

• the speech recognizer

• the user

The MARA hardware consists of a Multibus PC board, a backplane with a
connector, a BXC cable, a pre-amplifier, and a microphone. The software components
include:

• The PC board pvogram-maraS6.com

• The MAM Daemon-mara

• The Low Level Recognition command lihY^ry-libmara.a
■ The Standard Mbrdiry -libmara.a

• Support libraries for various applications-libmarawindow.a

The MARA system in the broadest sense is the combination of equipment and
programs that are referred to as the SRI 'Berkeley' board. (Kavaler, 1986)

D. THE SRI BERKELEY' BOARD

The speech recognition board, as its name implies, is a single circuit board. This
board is built with a multibus interface and is modified to be inserted directly into the
SUN Microsystems computer workstation. The speech board is divided into two
separate subsystems. The front-end subsystem manipulates the input into a form to be
analyzed by a comparator subsystem where the voice templates are stored.

1. Front End

The utterance, in the form of a frequency vs. time signal, enters thru a series
of 16 bandpass filters. The outputs are rectified and then low-passed filtered over a
period of time. The signal is then divided into 10 millisecond frames. Each frame ". . .
is the average voltage a speech signal has in several frequency bands. The system
computes speech frames at a rate of one hundred times a second." (Murveit, 19S6)
During the process of computing the frames it checks for whether or not a word is
really being spoken (referred to as endpoint detection). Assuming that a word is being
spoken, the system varies the spectral sampling rate dynamically. The spectral
difference of adjacent frames are then compared, and if the distance is insignificant
then the frame is discarded. This technique is called selective downsampling and it

30

reduces the data rate through the system, particularly the long steady-state sounds in
words. The result of disregarding the insignificant frames in this manner is improved
accuracy, real time vocabulary* processing, and expanded template storage memory.
The front end subsystem then downloads the frames into the comparator.
2. Comparator

As the name implies this subsystem compares the incoming frame with those
already in memory. This is accomplished by a technique called dynamic time warping.
The input frames are compared with the reference frames of the words in memory.
The sum of the differences of their spectral distances is computed. A score or cost for
each and even.' word in memory is then computed and the minimum value is sought.
The lower the score computed by the algorithm the better the recognition. As
discussed in Chapter 3, if the score is below a rejection threshold then the string
specified for the word is output. If the word score is above this value a non-
recognition occurs.

E. SUBJECTS

One civilian and one military* officer participated in the testing of the SRI speech
board. Both subjects were male 32 to 46 years old. The civilian (Ml) was very
experienced with many types and models of speech recognition systems, while the
military officer (M2) had less than 12 hours total exposure to speech systems.

F. TRAINING ALGORITHM

The training was conducted in a low noise speech lab at SRI utilizing a SHURE
SM-10 close-talking microphone. A training algorithm was used to develop the
templates for each speaker. This speaker dependent system requires the user using the
the training algorithm to specify how many training passes are desired as well as the
"cluster" size and method of input. This would allow one to input utterances from a
tape recording and have the algorithm form templates on a fixed number of passes
from the recording. The cluster size is an averaging technique which is the essential
ingredient in creating templates. To form a cluster, an initial template (the first
training pass usually) is compared against another utterance for that word or phrase.
The spectral distance is calculated and compared to the initial utterance(s) in memory.
If the minimum average distance is less than the distance specified in the algorithm,
then one template is formed. Otherwise the system will indicate that a template could
not be formed since the spectral clusters were outside the limits. The trainer program

31

then will prompt for more repetitions in an effort to generate a single template. If after
three more repetitions a single template still could not be created from the additional
utterances, two templates for the same word are computed. Each template and spoken
word is placed alphabetically in a Unix directory. The templates are indicated by file
type .// while the utterances are identified by a .uJ. For example if the word "advisory"
is spoken twice in creating one template one would find the files advisory. tl, advisory. u 1
and advisory.ul. This is unique to this system and the advantages of this scheme will be
evident later in this chapter.

G. THE VOCABULARY

Any vocabulary file can be created by specifying the word prompt followed by
two colons, then the keystrokes or output string. This file is in the working director}'
and is specified when invoking the trainer algorithm. In this particular experiment the
subjects used a 100 word initial vocabulary taken from the 1000 word set used by SRI
(Appendix A). A second vocabulary which was used in extensive studies conducted at
the Naval Postgraduate School (Poock, 1981, 1986a) was sent directly to the SUN
workstation at the host (SRI-BOZO) via the DDN. This vocabulary of 240 utterances
is shown on the data sheet in Appendix B. It is divided into five groups of words
based on the number of syllables. There were 10% one syllable words, 30% two
syllable words, 20% three syllable words, 20% four syllable words, and 20% five or
more syllable words. These words were selected from commands typically used in a
command center.

H. PROCEDURE AND DATA COLLECTED

Several different testing periods were scheduled over a three month period. Both
subjects traveled to the SRI International building in Palo Alto, Ca. to participate in
the testing. The session started by logging onto the SRI-BOZO net via the Sun
Microsystems Computer terminal. The appropriate windows were displayed and the
MARA system was automatically enabled during the login sequence.

The trainer program was used only once for each vocabulary. One user (MI)
used three training passes while the other user (M2) only used two passes. There was
no need throughout the three months to retrain the vocabularies. A selective retraining
of several words was accomplished to demonstrate the ease of retraining or adding new
words.

32

Under the main directory of XPS were the subdirectories of templates
POOCK.TEMPLATES and MIKE.TEMPLATES. The word recognition program was
enabled and the file of 100 words or 240 words was called. The program automatically
searched the alphabetical subdirectories and loaded the proper templates on to the
speech board. It took an average of 130 seconds to load the 240 word templates. For
data collection purposes each session was recorded to a file with the lowest five words
and their scores for each utterance. When possible- the other subject would record
errors as he witnessed them to confirm the recorded data. Additionally, any
abnormalities or peculiarities the system would display would be more apparent to the
observer and thus free the subject to concentrate on the word list.

In an effort to demonstrate the robustness of the system, the different lists were
read with varying speeds. The vocabulary was tested forward, backward, and randomly
at both a normal speaking rate and then at a significantly quicker pace. In addition.
the subjects attempted to demonstrate the interoperability of the same voice patterns
between the two subjects by using each others templates. A joint template was
attempted but due to the relatively small spectral distance allowed in the training
algorithm cluster averaging technique, after four passes no single joint template could
be created.

Several runs were conducted in a noisy environment. A cassette tape of
machinery noise was played at a level of 74 db(A) at the microphone. This level is
considerably higher than one could expect in a command and control environment
even in a shipboard tactical decision center.

The vocabulary can easily be modified by editing the file. If a file is modified to
include a word not yet trained, the speech program indicates that it could not find a
template for that word. Otherwise, it would load any template that was specified in the
vocabulary regardless of whether or not it was trained at the same time or a part of
another vocabulary.

During one of the testing periods, the subjects used a syniactic feedback system
demonstrated by SRI to NAVELEX in July 1934. (Murveit. 1986) The syntactic
feedback system is a specially designed algorithm to correct recognition errors in a
sentence. The grammar is structured as a finite state machine with beginning, end. and
transition states. The program would compute the least-cost path through a scries of
weighted arcs and then select the recognized sentence. For instance, in a data base
query if a word or words were misrecognized by the recognition system, it could be
corrected by the syntactic feedback algorithm.

33

Throughout the testing period it was evident that a good background in the
UNIX operating system and familiarity of the MARA system were major prerequisites
to effective use of the speech recognition system. Software improvements in user
interaction and a well written operating manual for reference would have been helpful.

I. RESULTS

1. Accuracy

Results for the 1000 word vocabulary tests conducted by SRI reported in
Murveit (19S6) are shown below in Table 1. Ml, M2, Fl, and F2 refer to individual
male and female subjects.. The percentages refer to word recognition.

TABLE 1

SRI 1000 WORD RECOGNITION PERFORMANCE

Ml

89-91 %

M2

89-93 %

Fl

91-93 %

F2

86-90 %

The data shown in Table 2 reproduced from Murveit (1986), reflect the results
of SRI's speech recognition system utilizing the TI-20-word data base used to test
commercial speech recognition systems. (Doddington and Schalk. 1981)

The results of the tests conducted by our subjects appear in Tables 3 through
6. These tables represent the trials with the variability in speech speed and no
maximum rejection threshold specified.

A two sample T test utilizing an Arcsin Transformation criteria was completed
using MI SI-TAB statistics package showing no significance between the two means of
our subjects at the 0.05 level of significance. (Minitab, 1981)
2. Interoperability of Voice Patterns for Different Users

The results of the interoperability tests are shown in Table 5 showing an
obvious decrease in accuracy. The computed scores or differences between the

34

TABLE 2

SRI TI DATA BASE PERFORMANCE (ERRORS OF 320)

16 SPEAKERS TOTAL 13 ERRORS

.25%

MEAN ERROR RATE

320 UTTERANCES

EACH

TABLE 3
NTS 100 WORD VOCABULARY TEST

Ml

94-98 %

8 TRIALS

AVG 96 %

M2

91-99 %

12 TRIALS

AVG 97 %

TABLE 4

NTS 240 WORD VOCABULARY TEST

Ml

95-100% 7 TRIALS AVG 97 %

M2

9S-100 % 7 TRIALS AVG 99 %

recognized words and the templates were on the average 10 points higher than the
mean of their scores with their own templates.

33

TABLE 5

INTEROPERABILITY TESTS

Ml using M2 Templates 80-89 %

3 TRIALS

M2 using M2 Templates 78-86 %

3 Trials

3. Accuracy in a Noisy Environment

The endpoint deieciion process which is computed in the front end section of
the card also keeps track of the background noise level and effectively ". . . eliminates
moderate room noises and maintains proper signal levels in the converter and analysis
circuits." (Murveit, 19S6) The background noise elimination features oi" the
microphone and the system allowed it to perform with virtually no degradation in
recognition performance. It is interesting to note that the system was not capable of
any recognition at approximately 76db(A). Table 6 shows the results in a noisy
environment.

Ml
M2

TABLE 6
NOISY ENVIRONMENT

99 %
96-98 %

2 TRIALS
2 TRIALS

J. SYNTACTIC FEEDBACK

The subjects during oi\g testing session exercised the syntactic feedback system
using a limited vocabulary and allowable sentence structure. There are a number of
questions which are suggested by Murveit (1986). These issues should be pursued,
since there is an increase in accuracy realized in using this algorithm.

36

K. CONCLUSIONS AND RECOMMENDATIONS

The purpose of these tests was to examine the voice recognition performance of
the SRI 'Berkeley' 1000-word discrete speech recognition board. The results of our
testing confirms the results reported by SRI Project 6096. (Murveit, 1986) Their
1000-word speech recognition system is very accurate and quite fast. Throughout the
entire study, no degradation of the templates occured. The experiment was conducted
entirely on initial templates. Despite the variability in speaking rate, three months of
broken testing, and testing in a noisy environment, the system performed proficiently.

However, the SRI 'Berkeley' board in its present configuration does not meet all
the requirements necessary to be a viable interface in the CCWS. In spite of
commercial discrete speech recognition system vendors advertising an input rate of 60
words minute, discrete speech recognition systems are not suitable for a Command and
Control environment. The user must modify his speaking rate by pausing alter each
utterance to effectively use the system. It would be insensitive to the ultimate users in
a CCWS environment to assume that discrete utterances in a high tempo, high
pressure, and possibly high threat situation is even remotely acceptable. A connected
or even a continuous speech recognition system is the only suitable alternative. This
gives the Commander the best opportunity to process information quickly and
accuratelv allowing him more time to enact a timelv and knowledgeable decision.

37

V. TEST, ANALYZE, AND EVALUATE A COMMERCIAL CONNECTED
VOICE RECOGNITION SYSTEM IN A WARGAMING ENVIRONMENT

The previous chapter analyzed the reliability of a 1000 word discrete speech
recognition system. The SRI speech board is a state-of-the-art system which was quite
good and very accurate. The disadvantage was, of course, utilizing a discrete system in
a command and control environment.

The purpose of this chapter is to analyze the performance oC a relatively
inexpensive, commercially available continuous speech system. The VOTAN 6050
Model II product was examined for its applicability and adaptability to a command
and control environment in a particular Naval Warfare Interactive Simulation System
(NYV15S). VOTAN has been used in many experiments, tests, and applications and is
regarded by many as a very capable speech recognizer. For example, in the Navy's air
traffic control trainer and simulator this same recognizer was demonstrated and
performed quite well. The VOTAN was used in this experiment to focus on four major
areas:

(1) An application of a continuous speech svstem in a Command and Control
environment similar to a workstation module.

(2) Investigate anv significant differences in the ability to input commands by
speech or keyboard entry.

(3) Investigate the possibility of utilizing a speech recognition svstem in Naw
Tactical trainers to overcome the dead time in learning the game command
keystrokes and entry procedures.

(4) Investigate anv significant differences in speed of command entrv for users
with familiarity with standard Navy phraseology versus those unfamiliar with
using speech recognition systems.

There is considerable time expended at every tactical trainer by the users in

familiarizing themselves with the equipment and game command entry procedures.

This 'dead' time could be eliminated by using a standardized vocabulary as used in

Navy contact reporting procedures and incorporating speech recognition to minimize

keyboard operation and special game commands. The result would be an increase in

useful tactical trainer time. Before examining the VOTAN speech system we will

briefly describe NWISS and the similarities to the proposed specifications for the

Command and Control Workstation (CCWS).

38

A. DESCRIPTION OF THE NAVAL WARFARE INTERACTIVE SIMULATION
SYSTEM (NWISS)

NWISS is a real-time, user-interactive simulation of naval warfare. Its mission
was originally to train senior Naval Officers in force-level tactical decision making and
management of command and control. The NWISS game resides on a VAX 11,780
computer, and a network of peripheral VT100/102, ADM31 terminals and RAMTEK
graphics terminals to provide the necessary displays and interactive stations. The
equipment is located in the Naval Postgraduate School Wargaming Analysis and
Research (WAR) Laboratory. There is a sufficient amount of equipment to support
three separate bays or areas to simulate disjoint command and control modules.

The equipment available in the wargaming and research laboratory is very similar
to the equipment for the CCWS. The Distributed Command System (previously
shown in Figure 1.1). shows the Interim Battle Group Tactical Trainer (IBGTT). which
is a component to be interfaced into the local area network. NWISS is to be
integrated into the IBGTT network in 1987. In applying a continuous speech
capability on the NWISS. we can analyze the requirements for a continuous speech
system in a C environment.

The RAMTEK monitor is the display system used in the NWISS modules. The

presentation is nothing more than a typical Naval Tactical Data System (NTDS)

picture with some exceptions and is similar to the display envisioned for the CCWS.

All ships, planes, and submarines are displayed utilizing standard Navy symbology as

shown in Figure 5.1, with some differences. The exceptions to standard shipboard

NTDS console display are summarized below:

NWISS has color enhanced symbology (An excellent screen improvement).

The track symbology in NWISS does not reflect engagement status of tracks.

Track information is available onlv on display boards and is not accessible from
the graphic display screen.

Electronic (ESM) and acoustic (SONAR) emissions lines of bearing are color
coded as well.

Old tracks change to yellow to indicate a fading track.

NWISS does not have representative svmboloev available in NTDS to indicate
type of platform.

NTDS has balltab capability for immediatelv obtaining information on the
status of tracks.

The color scheme displays all known friendly forces in blue, enemy forces in red, and

unknown contacts in white, with a fading tracks indicated in yellow.

39

COMMON (NTDS & NWISS)

AIR SURFACE SUBSURFACE
/\ <Q> \y HOSTILE

| | UNKNOWN

SPECIAL POINTS

CAP STA.

DATUM
ACOUSTIC FIX

r^

o

EWFIX

\^y FRIENDLY

NTDS

HOSTILE RAID MANY >/\ ASSIGNED SAM/GUNS

FEW >^ ASSIGNED CAP

ASSIGNED fj^ ASSIGNED

ENGAGED fT\

AMPLIFYING I.D

^G) © m rh rr\

CAP CV CG/CGN

T
X

FIXED

ASW

AIR

WING

HELO

TANKER

ASW

Figure 5.1 NTDS Symbology.

B. SCENARIO

The scenario for the NWISS game was designed to place subjects in situations
requiring the input of many combinations of the various commands available. It was
the first exposure for most of the subjects to a multi-threat Naval wargame since it was
the introductory simulation course for students of the Naval Postgraduate School
Command and Control curriculum. Each group of students embarked in separate
aircraft carriers or command and control modules. 1 lie objective was designed to
demonstrate:

• High Resolution Color Graphics

• Friendly man - machine interface

• The level of detail required to plan, run, summarize, and analyze a relatively
low level waruame

40

• The N'PS WAR Lab capabilities

Additionally, the purpose of each of the runs was to familiarize the subjects with
the game and experiment with the various commands and display boards. The actual
situation briefing used in these tests is included in Appendix C.

C. VOTAN SPEECH RECOGNITION SYSTEM MODEL 6050 SERIES II

The VOTAN VTR 6050 Series II is a stand alone unit which can interface with
any system supporting a standard RS-232 port. It has the ability to operate in two
distinct modes: Voice Terminal (VTR) and Voice Peripheral (VP). The VTR mode
allows the equipment to interface directly between a terminal and a host. This is the
mode that was used in the NWISS game with an ADM 3 1 terminal and the VAX
11 780 as host. The configuration to run NWISS with the VOTAN appears in Figure
5.2. The VP mode is designed for telephone-based applications. This mode was not
used in this experiment and will not be discussed.

CONFIGURATION TO RUN
NWISS WITH VOTAN

VAX 11/780

ADM31

J

/

*

V<
6

DTAr*
050 II

J

>

x

T

ERMINAL HOST

Figure 5.2 Configuration To Run NWISS With The VOTAN.

1. Vocabulary Size

The VOTAN 6050 Series II has three internal components which support its
vocabulary. These are:

• VTR System Memory (approximately 500K)

• Floppy Disk Memory (maximum of 76()K)

• Voice Card Memory (maximum o( 22K)

41

In addition to these components there is also the possibility of storing voice
files on the host computer. This was not used since the vocabulary was small enough
to be stored directly in system memory. The average word or template uses 200-250
bytes of memory. When the system is fully loaded, there can be 2000-3000 words in
main memory. It is important to note that all voice recognition takes place on the
voice card. The voice card can accommodate up to 50 words (from the 2000-3000 in
main memory') at a time. A tradeoff can be seen in the number of words vs. the
number of templates for each word. The more accuracy required, the more templates
needed for each word, and the fewer words loaded into each active set.

The main memory can contain multiple sets and takes only about 150 msec to
upload sets onto the voice card memory7. This can be done by tailoring the vocabulary
to switch automatically upon hearing a switch word or can be automatically switched
when a certain number of word(s) are recognized from an on-line set. A switch is a
mnemonic that is spoken by the user to load the voice card with a specific set of
templates. This file is transferred at a rate of 9600 baud. During the upload period the
VTR is automatically recording speech (up to 7 sees) to be searched immediately upon
completion of the swap. It is extremely fast and is virtually unnoticed by the user. It
is recommended in VOTAN Guide To Procedures, that one should limit the number of
words in a set to about 10 to 20. A set of this size will optimize recognition and
provide a quicker system response time.
2. Programming

The VTR 6050 Series II can be easily programmed. The key element in
optimizing the performance of the system is careful construction of the vocabulary so
as not to exceed the voice card memory limitations and to minimize set changes. With
the VTR in the off-line mode, (which blocks any keystrokes from going to the host), a
vocabulary is entered directly onto the screen in an editor mode. The user specifies the
file name and then begins entering headings for the word sets followed by the actual
words in the set. The following is an example of a small file which is included to show
the various programming commands available: (VOTAN GTP and UG, 1985)

EDT NUMBERS -(this allows you to enter the EDITOR)

"(mode)
S-NUMBERS, *(this specifies the set name NUMBERS)

NS = COLORS, -(this is the pointer to the NEXT SET:)

*(COLORS which is)

42

CT= 2, *( automatically loaded after 2)

precognitions of this set)
CM ^indicates NUMBERS is a COMMON)

*(word always in memory)
ONE,HS = 1 "(ONE is the prompt and 1 is the)

*( string sent to the host)
TWO,TS=2 *(TWO is the prompt and 2 is the)

*( string to the terminal)
THREE.TS = 3\20 *(the \20 is the hexadecimal string)

*(for space to be)

*(sent to the terminal after the 3)
FOUR,HS = 4 -(FOUR is the prompt and 4 is the)

"(string to the host)

Appendix D is the listing of the vocabulary used for the NWISS game and will be
discussed later in the chapter.
3. Operation

While the VOTAN 6050 Series II is still in the off-line mode, the user's
vocabulary and templates are placed into memory. In addition to the set in memory
there are certain words called TASK WORDS which control operation of the VTR
when it is on-line, and a collection of words in the user tailored vocabulary which can
be indicated as COMMON words that are also a part of the total allowed templates on
the voice card memory. The user can specify an initial word set that will be activated
each time the system is initialized. Additionally, the user can specify whether or not
data buffering should be used. Data buffering allows the system to store a
predetermined number of strings or characters before outputting them to the host.
Data buffering can be extremely beneficial when a user needs to verify a string of
words prior to being sent. Numerous military situations require validation of codes or
strings to ensure proper actions upon receipt. The default condition is immediate
action when the word or phrase is recognized. These are some one time preliminary
set-up inputs. Once this is accomplished the system is ready to be put in the on-line
(ONL) mode. This sends the host string directly to the computer upon recognition.
These keystrokes are then returned by the host and displayed on the screen. The
keyboard can still be used and the VOTAN is transparent to the user when passing
these kevstrokes directlv to the host.

43

4. Training Algorithms

The VOTAN 6050 Series II offers two types of training algorithms:
single/discrete training and continuous training. In the single training mode, one
template is formed after each utterance. The continuous training method extracts
templates from a series of passes for each word in the set. This takes into account the
coarticulation of a word at the beginning, middle, and the end of a group of words.
Prior to entering the continuous training mode, the user must have at least two single
trained templates available for template extraction to occur. The user specifies the set
which he would like for continuous training. The algorithm then automatically selects
up to ten words at a time and presents to the user a series of five of these words in
random order on the screen. The user repeats all five words in a continuous manner.
It will then display two columns of words if a sufficient number of words were
recognized. The first column lists the words that were displayed as the prompts. '1 he
second column contains the words that the system recognized.

Several misrecognitions may be observed; however, the algorithm uses the
other correctly recognized words for forming the extracted templates. This ability to
develop these extracted templates enables VOTAN to make the claim of having a
continuous recognizer. The operator can manipulate the presentation during
continuous training to ascertain the progress of completion of a recognition matrix for
the current set of words being trained. The matrix has three columns for each word
indicating where the word occured in a string of words (i.e.. beginning, middle, or end).
There are some training passes where there will be an insufficient number of words
recognized and the system will prompt the user to continue training a new set. After a
certain number of passes or when the matrix is completely filled, the program will
terminate the training of that word group and continue with the next set often words.

Prior to operating the system in VTR mode which transmits the output strings
to the host or terminal, the user can invoke a program to test his templates and to
ensure voice card storage has not been exceeded. The output display upon recognition
consists of the recognized prompt characters and the recognition score. The
recognition score is computed from the spectral distances between the template and the
spoker word. Like the SRI system the lowest score is the best recognized word. The
recognizer has a minimum recognition threshold default of 50, but the user can modify
this value if desired. This level appears to be quite adequate for most applications.

44

D. SUBJECTS

Six male officers participated in this experiment. Five were Naval Officers from
various communities. Three had previous experience with the modeled systems and
were familiar with the terminology of giving similar orders. These were the individuals
used in validating the area of familiarity with battle group phraseology vs. having no
experience. All but one of the officers had less than 12 hours total exposure to voice
recognition systems. The other officer had about 100 hours experience with various
voice systems.

E. THE VOCABULARY

The vocabulary for the NWISS wargame consists of two major groups of
commands: DISPLAY and ACTIVE. The DISPLAY commands control all aspects of
the graphic plot as displayed on the RAM TEK monitor. The 'active' commands
consist of many different orders that could be given to ships, submarines, and aircraft.
There are actually a total of 230 allowed words that are recognized by the NWISS
game. The NWISS game requires that the commands be ordered in a particular way.
For example, after activate, the game would expect to see 7 different commands, and
would disallow other inputs. These same words could appear in different positions in
different correct commands to the host (this plurality in commands occurs throughout
the vocabulary). In addition, the number of options after identifying a force name can
range up to 50-60 possible commands, greatly exceeding the limitation of the voice
card. This peculiarity required a more general tailoring of the vocabulary7 to model the
NWISS word structure, since one could not tailor the vocabulary into finite sets
allowing only a small number of words to follow other words. It is a similar problem
experienced by SRI in formulating the valid structures used in formulating the finite
states used in the syntactic feedback system. Consequently, this made it impossible to
formulate the vocabulary- within the memory and template limitations without multiple
switch words.

Appendix D is the listing of the vocabulary used in this experiment. Note that
there are six major vocabularies or sets: Display, Ships, Commands to Units.
Numbers, Aviation, and Load. This was done to minimize the number of switches
necessary for full use of the commands. For example, an actual voice command for
activating an air search radar utilizing the VOTAN would be:

SHIPS SPRUANCE ACTIVATE AIR NUMBERS 1245 ENTER. (6 sees)

45

The bold words are the switch words for the two sets.
The same command by keyboard entry is:
FOR SPRUA ACTIVATE AIR 1245 <cr>
(28 keystrokes) (~ 10 sec) (NWISS, 19S3)

F. PROCEDURE

The training was conducted in the C WAR Lab at the same input terminal to
be used for the game. A SHU RE SM-10 close talking microphone was used for the
training and game play. The subjects used in the experiment were trained in individual
sessions on the VOTAN speech recognizer. The training took place in one session
which averaged approximately 75 minutes. The enrollment started by loading copies of
the commands as shown in Appendix D in active memory without any templates. An
overview of how the training was to be conducted was given including proper
microphone placement and description of the vocabularies.

Each subject started by generating two single trained templates for the set of
NUMBERS, (this set included all numbers 0-9 and letters A-Z). The set NUMBERS
was anticipated to require continuous training because of the extensive use of alpha-
numerics in commands. Following the individual training of this set, the continuous
training algorithm was invoked. Displaying the continuous training matrix during
training led to the discovery that the algorithm is not sophisticated enough to
determine exactly what order it should present the group if there are only a few unfilled
blocks left in the matrix. This can be time consuming especially if the processor is
experiencing some difficulties in developing an extraction template for a particular
word. Upon completion of continuous training there were now five templates for each
word in the set. It became apparent that this number would far exceed the number
allowed on the speech card and therefore all single templates were erased. The
remaining words were presented for two sets of single1 discrete training passes.

After all word sets were trained, each set was displayed with the total number of
templates and memory used. Task words' and 'common' words reside on the voice
card at all times. In all cases, three of the six possible sets had exceeded usable
memory, as shown in Table 7.

A review of the vocabulary and sets showed that 28 words were duplicated
intentionally in the composition of the sets. This design redundancy was to reduce the

46

TABLE 7
INITIAL TEMPLATE LISTING

VOCABULARY
SET

MEMORY

(BYTES)

AVERAGE

TASK_WORDS

1651

COMMON

2916

NUMBERS*

1S740

COMMANDS JTOJJNITS

23675

AVIATION

25382

SHIPS

S900

LOAD

16229

AVERAGE
NUMBER OF
TEMPLATES

8

IS

1 jj

148

142

50

86

SINGLE TRAINED TEMPLATES NOT INCLUDED

EXCEEDS VOICE CARD LIMITATIONS
(COMMON AND TASK_WORDS INCLUDED)

number of switches needed for the formulation of proper commands. Consequently,
there were actually four separately trained templates for these words in storage. Two
of these templates for these words were deleted from the active sets. In every case, an
average of 45 additional templates were deleted to bring the memory and number of
templates allowed within limits. The words that were reduced to only one template
were those words with many syllables and that were readily recognized. The actual
number removed varied according to the user and the way each word was enunciated.
That is. if utterances were fairly slow, more memory was required. Table 8 depicts the
average final number of templates and memory remaining in the actual individual files
for all users. The final test was to invoke the trainer program and ensure there were no
memory overflow or template overflow errors produced as the different sets were
loaded onto the voice card. It is recognized that having to delete templates causes a
corresponding decrease in recognition and is a significant limitation imposed by the
svstem.

47

TABLE 8
REVISED TEMPLATE LISTING

VOCABULARY
SET

MEMORY

(BYTES)

AVERAGE

TASK_WORDS

1651

COMMON

2916

NUMBERS

13761

COMMANDS_TO_UNTTS

16953

AVIATION

17304

SHIPS

8900

LOAD

15S54

AVERAGE
NUMBER OF
TEMPLATES

18

100

104

95

50

85

Each subject had no further training. At the start of the -game the subject's
revised templates were loaded into the recognizer. They were allowed to perform their
roles by inputting commands as necessary.

The short time available to conduct the tests precluded evaluating the
interoperability of data sets (i.e., one user operating from another's voice templates).
Although the system was not designed to accomplish this, it is a point of interest when
evaluating systems in a command and control environment. The purpose is that in the

event of a mishap to the active operator a slow transition to another operator would

2
have a negative impact on the C center operation. The time to exchange vocabularies

from one user to another was 62 seconds.

The level of noise in the module was not measured, but during the conduct of the
exercise the noise in the groups during discussions and administration was very similar
to those encountered in a real command and control center. The VOTAN gain can be
easily adjusted if necessary.

Additionally, the 240 word vocabulary (Appendix B) was loaded into the
VOTAN. A comparison of speech recognition accuracy of the VOATAN vs. SRI is
shown in Table 9 using subject M2 from the previous tests. The 240 word vocabulary

48

was loaded into 5 sets and with an average number of 96 templates and 19575 bytes of
memory per set to simulate the conditions present for the NWISS vocabulary'. It is
evident from the data that exceeding the manufacturers recommendations of loading
does in fact effect performance.

TABLE 9
SRI VS VOTAN 240 WORD RECOGNITION ACCURACY TEST

M2 SRI 99 % VOTAN 97.4 %

G. RESULTS

The experiment set out to focus on four separate areas:

(1) Demonstrate an application of a continuous speech system.

(2) Investigate any significant differences in the ability to input commands by
speech "or keyboard entry.

(3) Investigate the possibility of utilizing a speech recognition system in Navy
Tactical trainers to overcome the dead time in learning the game command
keystrokes and entry procedures.

(4) Investigate anv significant differences in speed of command entry of users
familiar with standard Navy phraseology versus those unfamiliar with using
speech recognition systems. '

The results from the three separate runs and data collected with the constraints

described show that the VOTAN in its present configuration was unable to adapt to

this C*" environment. This is primarily due to the limitations ol storage and processing

power of the voice card. The NWISS vocabulary is not suited for designing a distinct

branching method of words from one set to other sets for correct formulation of

commands. This inability to establish a tree architecture for correct command

structure, resulted in the number of words in most sets exceeding the recommended

number by 3.5 times. As discussed in the technical documentation and discussed

earlier, the optimum number of 10-20 words would increase recognition and provide a

quicker response time. With an average number of 55 words, the reaction time was

inordinately slow and misrecognitions were higher than expected. Speed of speech

input as stated by Kavaler (1986). is a function of:

• Speech rate

49

• The processing power of the speech recognizer

• ' The constraints placed on the way the user must speak (i.e., discrete vs.

connected, number of 'switch words'J.

Subjects entering commands by voice with these constraints were confused and
frustrated since the time delay for the recognition to appear on the terminal was often
slower than one would expect for keyboard entry. Likewise, if a misrecognition
occured at some point in the string a user would have to attempt to back out the
command or cancel it and start the entire entry over again.

The design of NWISS command entry procedures has some unique human
engineering advantages for keyboard entry. The host would not allow a command to
be entered if it did not form a correct entry. The terminal would beep and inhibit any
incorrect keystrokes. The user could type a question mark '?' and the list of acceptable
entries would be listed. Even though this occurred in the voice entry procedure as well
the user would be disappointed by the misrecognition and often forget the voice
command 'help' which would output '?'. Eventually, he felt more hostility and mistrust
toward the recognizer and got flustered, forgetting which set he was in and eventually
cancelling the entire command again.

The frustration from a misrecognition was also attributable to the unfamiliarity
with words in the sets and the proper NWISS command structure. The user usually
blamed his uncertainty in the set and command structure on himself adding to more
disappointment and disillusionment with the recognizer. In later trials, a combination
of voice and keyboard was used by some subjects. They used voice for certain words
and commands they felt comfortable with and then used the keyboard for the
unfamiliar commands Or for entries they felt required immediate and correct entry.

There could not be any determination o( advantages in utilizing a speech
recognition system in Navy tactical trainers to overcome the dead time in learning the
game command keystrokes and entry procedures. The human engineering in the design
of this particular wargaming system was extremely helpful both in providing assistance
and prompts, as well as accepting as few as four keystrokes for certain commands.
Further study is required in this area.

The subjects with some familiarity with wargaming had a distinct advantage over
those who did not, both with and without voice entry. This advantage could not be
directly attributed to the voice recognition application but was quite evident in the
level of play. They were more comfortable at the input terminal and were relied upon
by the other members in the group for advice to interpret the displays.

50

H. CONCLUSIONS AND RECOMMENDATIONS

Even though initially the VOTAN seemed very promising and an excellent
candidate for a C environment, this speech recognizer is not well suited for CCWS. It
failed because the vocabulary limits of the voice card and the processing power of this
recognizer were exceeded by the demands of the NWISS vocabulary. Consequently,
the recognition and output speed were jeopardized. The large 1000 word vocabulary
and real-time processing is necessary in the CCWS application for data base queries.
Additionally, the user is required to memorize which set is active and the 'switch'
words needed to enter the various sets. The user using the VOTAN must adapt his
speech to the recognizer which is unacceptable. The recognizer must be an extension
of the commander not a hinderance.

The combination voice and keyboard entry employed by some o[ our subjects
during the end of the testing indicates a possible area for future study. The application
of speech entry in conjunction with keyboard, mouse, or balltab manipulation should
be investigated. The balltab is the exclusive device for an NTDS console in a
shipboard command and control center. This would allow a smaller, more tailored
vocabulary integrated into existing systems to aid the user, particularly if that
individual must be positioned at a console or terminal.

51

VI. CONCLUSIONS

It is intuitive that the commander who can manage and process the tremendous
flow of battle information the fastest will have more time to determine a response or
make decisions which are always ahead of his adversary. As the dependency of the
commander on computing resources increases, it is only natural to expect greater
demands upon the man-machine interface. By including a speech recognition system
on the CCWS, the commander would realize a faster information processing rate. This
would result in the commander acquiring more knowledge in a faster time on which to
base his decision. As Sun Tzu, the famous Chou Dynasty philosopher and military
strategist once stated ". . . knowledge is power and permits the wise to conquer without
bloodshed and to accomplish deeds surpassing all others."

This thesis evaluated the performance of a state-of-the-art 1000 word discrete
template matching system and a commercially available VOTAN continuous speech
recognition system. The requirements specified for the CCWS were:

Large vocabulary (capacity > 1000 words).

Real-time response.

Very high recognition accuracy ( > 98%).

Adaptable to the user, (i.e., the user should not have to modify or alter his
speaking rate significantly)

No deterioration in accuracy in noisy and stressful environments.

The systems evaluated in this thesis did not fulfill all the requirements for the speech

application in the CCWS. Each system had its advantages and disadvantages which

were discussed in the conclusion of each respective chapter. Currently, there is not a

system commercially available capable of meeting all these requirements.

Even though neither system met all the requirements established for the CCWS,
recent literature reflects the improvements in the Strategic Computing Program, in
particular, phonetic recognition. Speech systems capable of meeting and exceeding
these specifications are not far away. In fact, CINCPACFLT is scheduled to test and
evaluate the speech recognition system being developed by the Strategic Computing
Program. (Strategic Computing, 1985)

As computers become more and more capable of displaying, storing, and
processing information, it is only natural to assume that the interface between the user

52

and computer should be. optimized. We all can recount from our own experiences, ". .
. the costs of poorly designed interfaces. Coming in many forms, the cost can include
degraded user productivity, user frustration, increased training costs, and the need to
redesign . . ." (Foley, 1984). For these very reasons, the design of even- interface for
an interactive user-computer must be of utmost importance. Speech recognition has
long been thought of as the ideal interface and must be considered for all future
systems.

53

APPENDIX A
SRI 100 WORD VOCABULARY

a

dinner

manner

rose

able

direction

many

round

aboard

discovered

March

run

about

distance

mark

running

above

do

market

said

accept

for

Mary

steps

according

foce

material

still

account

forced

matter

stock

across

foreign

may

stone

act

forget

maybe

stop

both

form

our

stopped

bottom

forty

out

store

box

forward

outside

story

break

found

over

straight

bring

for

own

street

broken

I'd

page

U.S.

brought

I'll

paid

under

Brown

I'm

paper

understand

building

I've

Paris

union

built

idea

part

university

development

ideas

right-paren

unless

did

if

river

until

didn't

immediately

road

up

different

important

Robert

difficult

impossible

room

54

APPENDIX B
240 WORD VOCABULARY

one

yankee

Gary_Poock

carriage_return

Iran

S wee den

login_Poock

accat_title

load_gld3

Poock_NPS_password

three

logout

red_sphere

zero

November

use_that_one

Captain_Ebbert

up_in_detail

level_two_viewer

genisco_zero_parameters

five

alpha

charlie

echo

juliett

move_it_left

San_Francisco

engineering

voice_technology

Russian_version_of_Hormuz

eieht

two

air_routes

load_the_gun

load_the_server

Japan

Europe

level_two

strait_of_Hormuz

connect_to_charlie

change_directory_to_hunter

four

graphics

steam_plant

seven

move_it_down

spirograph

close_out_charlie

United_States

North_Atlantic_.\lap

M editerranean_Chart

six

bravo

delta

foxtrot

romeo

sierra

application

human_factors

central_expressway

filc_transfer_protocol

nine

55

hotel

kilo

oscar

move_it_right

Vietnam

advisory

business_meeting

speech_recognition

efficient_transmission

golf

quebec

victor

xray

move_it_up

Tokyo

down_in_detail

criteria

suitability

identification

course

command

bingo

proceed

altitude

relocate

available

track esm

command_and_control

enemy_detection

launch

cancel

bearing

orders

satellite

negative

india

lima

pappa

uniform

Korea

interactive

continuous

continuous_speech

system_integration

mike

tango

whiskey

zulu

Bangladesh

Hollister

corporation

advantages

radiology

automatic_recognition

speed

attack

report

station

recover

designate

plot_esm

designate_track

probability

probability_of_detection

fire

message

label

copy

envelope

correlate

56

combination

maneuver_delay

Task_Force_Commander

proceed_to_New_Delhi

time

surface

minefield

shore_based

execute

enemy

Connecticut

Oklahoma

California

place_a_marker_on_Paris

bingo_all_craft_immediately

neutral

sensor

Stockton

air_field_name

track_friendly

bearing_and_distance

Minnesota

Eisenhower

relocate_the_Sunfish

take

Georgia

Texas

Utah

latitude

Ohio

flight_controller

Pango_Pango

lay_a_barrier

attack_barrier_target

scope

sensor_delay

Alabama

North_Carolina

place_a_circle_on_M oscow

shoot

refuel

distance

contact

submarine

order_name

Indiana

Pennsylvania

South_Dakota

map

grid

missile

Adak

New_York

track_unknown

track_neutral

Louisiana

Colorado

Ne\v_Mexico

refuel_the_Connie

place

Vermont

Daniels

platform

longitude

torpedo

Trans_World_Airlines

keep_on_station

ground_control_approach

Atlantic_Data_Base

drop

57

Bangkok

Brisbane

Antwerp

Arkansas

user's_guide

Acapulco

Yokohama

Diego_Garcia

Pacific_Data_Base

Maine

Portland

Aspro

red_fox

blue_force_one

Baltimore

Sevastopol

chronometer

plot_all_submarines

Iberian Carrier

Bombay
Canton
Africa
Saigon
Kitty_Hawk
Vladivostok
Sea_of_Japan
Indonesia
Arabian_Tanker
save

Rangoon
Kiev-
Naples
Calcutta
Wyoming
Honolulu
John_Kennedy
United_Air_Lines
West_German_Torpedo

58

APPENDIX C
SCENARIO BRIEFING

FROM: COMSEVENTH FLEET

TO: COMMANDER. TASK GROUP ONE PT ONE

COMMANDER, TASK GROUP ONE PT TWO

OPORDER 00003

1 THIS MESSAGE CONSTITUTES AN OPERATION ORDER FOR CTG
ONE PT ONE AND ONE PT TWO. IT CONSISTS OF GEOPOLITICAL
BACKGROUND. COMPELLING EXECUTION OF OPERATION. TASK
FORCE ORGANIZATION. OPERATION OBJECTIVES. SUMMARY OF
OPPOSING FORCES. AND DIRECTION CONCERNING CONDUCT OF
OPERATION.

2 DURING THE LAST 48 HOURS THE CVBGS HAVE DRAWN NEAR TO
EACH OTHER AND NOW MAY BE ORGANIZED INTO A TASK FORCE
OF CONSIDERABLE SIZE. AS LIGHT DAWNS THE JFK HAS
RECOVERED THE LATE NIGHT LAUNCH WHICH WAS CYCLIC DUE
TO THAT CARRIERS CLOSER PASSAGE TO ENEMY LAND BASES
AND DUE TO THAT THE JAPANESE ISLANDS THAT COULD NOT BE
ASSUMED TO BE FRIENDLY. THE AIR COMPLEMENT FIAS BEEN AT
WORK FOR AT LEAST 48 HOURS. KITTY ON THE OTHER HAND
HAS JUST LAUNCHED A CAP GRID WHICH IS PROCEEDING TO
POSITION. IT INCLUDES AN E2 AND AN S3.

A. AN E3A (AWACS) WAS SUPPOSED TO ARRIVE ON STATION OUT
OF ADAK ON AN AIR FORCE MISSION ABOUT ONE HOUR AGO.
HOWEVER SHE HAD NO REPORTING RESPONSIBILITY TO THE OTC AND
HER PRESENCE HAS NOT AS YET BEEN CONFIRMED. P3S ARE
DEPLOYED IN SUPPORT HOWEVER. .

TASK GROUP ONE PT ONE CONSISTS OF THE FOLLOWING SHIPS
LOCATED 12 HOURS PRIOR TO THE START OF YOUR RUN FOR RECORD
AS FOLS:

59

USS KITTYHAWK 46-30N/157E (APPROX)

USS WICHITA

USS KNOX

USS SPRUANCE

USS RATHBURNE

USS WILSON

USS MCCORMICK

USS FOX

USS LOS ANGELES

USS OMAHA SOJ

PATRON FOUR SIX IN PLACE MISAWA AB. 40-00N 141-50E.

PATRON SEVENTEEN IN PLACE, ADAK AB, 51-50N 176-30W.

UNSUBORDINATED AWACS DET IN PLACE, ADAK.

CVBG 1.2, JFK TASK GROUP CONSISTS OF THE FOLLOWING UNITS:

USS JOHN F. KENNEDY 46-30N 155E (APPROX)

USS IOWA

USS LONG BEACH

USS JOHN ROGERS

USS TURNER JOY

USS JOHN HANCOCK

USS MAC

USS FURER

USS GAR (NEW CONSTRUCTION SSN) SOJ

4. OPERATIONAL OBJECTIVES: (A REPEAT)

THE SEA OF OKHOTSK AND THE BASES WHICH SURROUND IT PROVIDE
A PRIMARY SANCTUARY FOR THE SOVIET FAR EASTERN FLEET.
PROCEED TO A POSITION FORM WHICH YOUR COMBINED FORCES CAN
INTERDICT SURFACE AND SUBSURFACE FORCES AND LAUNCH STRIKES
AGAINST THE SOVIET LAND BASED AIR STRONGHOLDS. PREPARE TO
FIGHT YOUR WAY IN AND STAY AS LONG AS POSSIBLE.

60

• PRIMARY MISSION ONE

PLAN FOR AND BE PREPARED TO CONDUCT A PREEMPTIVE AIR RAID
ON PETRO WHEN IN POSITION AND WHEN DIRECTED BY HIGHER
AUTHORITY.

• PRIMARY MISSION TWO

SEARCH FOR. IDENTIFY AND REPORT, THE SOVIET MINSK BG, AND ANY
RED SUBMARINES WHICH MAY BE ENCOUNTERED. BE PREPARED TO
CONDUCT SHORT NOTICE PREEMPTIVE ATTACK ON THESE FORCES
WHEN DIRECTED BY HIGHER AUTHORITY.

5. SUMMARY OF OPPOSING FORCES: ANTICIPATED OPPOSING FORCES
CONSIST OF THE SOVIET TASK GROUP COMPRISED OF:

ONE MINSK CLASS CGH

ONE KASHIN CLASS CGL

ONE KREST II CLASS CG

TWO VICTOR CLASS SSN

TWO CHARLIE CLASS SSGN

ONE ECH02 CLASS SSGN

INTELLIGENCE SOURCES INDICATE POSSIBILITY THAT

ADDITIONAL SURFACE UNITS OF UNKNOWN TYPE MAY HAVE

DEPARTED VLADIVOSTOK WITHIN THE PAST 36 HOURS.

ALTHOUGH THIS IS , AS YET, UNCONFIRMED.

24 HOURS PRIOR TO THE START OF YOUR RUN FOR RECORD, THE
SURFACE FORCES WERE IN THE SEA OF OKHOTSK. IT IS ANTICIPATED

THAT ONE SUB WILL CONTINUE WITH THE SOVIET BG DURING THE

LAST 36 HOURS ONE HOSTILE SSN HAS BEEN DETECTED IN THE
VICINITY OF KITTY. EVASIVE ACTION AND BEST SPEED MAY HAVE
LEFT IT BEHIND FOR THE TIME BEING, HOWEVER. SPEED OF TASK
GROUP ADVANCE HAS BEEN SLOWED AND VIGILANCE TO THE REAR IS
ADVISED. THE CONTACT THOUGHT TO BE SHADOWING THE JFK WAS
NEVER CONFIRMED BY CVBG FORCES OR THE FURER ON HER TRIP

61

NORTH. THE REMAINING SUBS ARE EXPECTED TO BE IN POSITION TO
OPPOSE YOUR TRANSIT NEAR THE ISLAND PASSAGES NOTHEAST OF
HOKKAIDO. INTEL STILL ESTIMATES THE GREATEST THREAT WILL BE
FROM (1) LAND BASED AIR OF REGIMENTAL SIZE GROUPINGS. AND (2)
FROM SSNs THAT ARE CURRENTLY DEPLOYED OR WILL DEPLOY
SHORLY. THE SOVIET TASK GROUP CAN BE EXPECTED TO OPPOSE
ENTRY TO THE SEA OF O TO SOME DEGREE.

6. DIRECTION CONCERNING THE CONDUCT OF THE OPERATION: THE
CONDUCT OF THE OPERATION IS AT THE DISCRETION OF THE OFFICER
IN TACTICAL COMMAND WITHIN THE FOLLOWING CONSTRAINTS AND
POLICY GUIDANCE:

1 DEFCON CONDITION TWO. WE ARE NOT AT WAR. IF POSSIBLE.
AVOID ACTIONS WHICH COULD PROVOKE A WAR. CONFIRM AS
EARLY AS POSSIBLE WHICH COMMANDER CVBG 1.1 OR CVBG 1.2.
WILL BE OTC. KITTY IS STILL THE ONLY SHIP WITH KEYING
MATERIAL NECESSARY TO GAIN LAND BASED AIR SUPPORT
FROM ADAK (THIRD FLEET) AND MISAWA (SEVENTH FLEET).
EXPECT LATE BREAKING GUIDANCE FROM THIS HEADQUARTERS
AS EVENTS IN EUROPE COULD SIGNAL THE START OF ACTIONS IN
THIS THEATRE.

2 WEAPONS ARE TIGHT AT THIS TIME. WEAPONS FREE STATUS
MUST BE REQUESTED FROM ORIG UNLESS ATTACKED, IN WHICH
CASE RESPONSE IN KIND ONLY IS AUTHORIZED. THAT IS TO SAY
THAT THE LOSS OF AN AIRCRAFT MAY NOT BE RESPONDED TO
BY AN ATTACK ON A SHIP. MINIMIZE ESCALATING ACTIONS.

• THE FIRST CHALLENGE WILL BE TO ORGANIZE THE COMBINED
TASK GROUP INTO AN EFFICIENT FIGHTING UNIT. NOTIFY THIS
HEADQUARTERS OF ALL SIGNIFICANT DECISIONS. YOUR PLAN
OF OPERATIONS, IN BRIEF, IS OF PRIMARY INTEREST.

• TO ENSURE SUSTAINABILITY IN THE EVENT OF A PROTRACTED
CAMPAIGN ONLY 36 AIRCRAFT MAY BE AIRBORNE AT ANY GIVEN
TIME FROM EACH CARRIER (TOTAL OF 72). THIS DOES NOT
INCLUDE LAND BASED P3s OR AWACS AC UNDER THE CONTROL

62

OF THE CARRIER. PERMISSION TO USE THIRD FLEET ASSESTS
MUST BE GAINED FROM THIRD FLEET. VIA SEVENTH FLEET.
PRIOR TO ISSUING A LAUNCH COMMAND.

SUBMIT YOUR PLAN OF ACTION PRIOR TO THE RUN FOR RECORD
CONTAINING:

1 BELIEVED ENEMY INTENTIONS:

2 YOUR INTENTIONS:

3 CONTINGENCY PLANS:

63

APPENDIX D
VOTAN VOCABULARY FOR NWISS

This file is the vocabularies set up for the interactive battle group game in the
war lab.
COMMON WORDS SET

001 COMMON
WRONG
ENTER
HELP
DISPLAY

COMMANDS JTO_UNITS
NUMBERS
AVIATION

SHIPS
LOAD

TASK WORDS SET

002 TASK_WORDS
GO_TO_SLEEP
LISTEN_TO_ME
INITIALIZE
VERIFY

003 WRONG,HS = \0B,CM

004 ENTER,HS = \OD,CM

006 HELP,HS = ?,CM
DISPLA Y WORDS SET

008 DISPLAY,CM

CANCEL, HS = CANCEL\20
CIRCLE,HS = CIRCLE 20
GRID,HS = GRID\20

RADIUS,HS= RADIUS 20
SHIFT,HS = SHIFT\20

DESIGNATE,! IS = DESIGNATE\20

64

XM ARK.HS = XMARK\20

CENTER,HS = CENTER ,20

FORCE,HS = FORCE\20

POSITION.HS = POSITION 20

DROP,HS = DROP 20

ERASE, HS = ERASE, 20

ESM.HS=ESM,20

PLOT.HS = PLOT 20

LINE_OF_BEARING_SONAR.HS= LOB_SONAR 20

LINE_OF_BEARING_ESM.HS = LOB_ESM\20

COMMANDS TO UMTS SET

009 COMMANDS_TO_UNTTS.CM
TIME.HS = TIME 20
AIR.HS = AIR 20
RADAR,HS=RADAR\20
EMITTER,HS = EM ITTER 20
ALTITUDE.HS = ALTITUDEl20
BEARING.HS = BEARING\20
POSITION. HS= POSITION 20
BLIP_ON,HS = BLIP ON\20
COURSE. HS = COURSE\ 20
OFF,HS = OFF\20

DESIGNATE. HS= DESIGNATE 20
FRIENDLY,HS= FRIENDLY 20
UNKNOWN. LIS = UNKNOWN 20
EXECUTE, HS= EXECUTE 20
LAUNCH. HS = LAUNCH 20
PERISCOPE.HS = PERISCOPE 20
CHAFF,HS=RBOC 20
SUBMARINE. HS = SUBMARINE 20
HANDOVER.HS = HANDOVER 20
JOIN.HS = JOIN 20
RECOVER,HS= RECOVER 20
SEARCH. HS = SEARCH 20

BEARING.HS = BEARING 20
BACKSPACE,HS = \08
SPACE,HS = \20
TR.ACK.HS = TR.ACK 20
OLD.HS=OLD,20
SONAR.HS = SONAR 20
PLACE A,HS= PLACE 20

ACTIVATE. HS = ACTIVATE 20
SURFACE. HS= SURFACE 20
ESM.HS=ESM 20
SONAR,HS= SONAR 20
BARR1ER,HS= BARRIER 20
FORCE. HS=FORCE\20
TRACK. HS = TRACK' 20
BLIP_OFF,HS = BLIP OFF\20
COVER,HS = COVER 20
DEPTH, HS = DEPTH 20
ENEMV,HS= ENEMY 20
NEUTRAL.HS = NEUTRAL 20
EMCON.HS = EMCON 20
FIRE.HS=FIRE 20
ORDERS. HS = ORDERS 20
PROCEED. HS= PROCEED 20
REFUEL. I IS = REFUEL 20
CEASE. HS = CEASE 20
INFORM, HS= INFORM 20
RECALL. HS= RECALL 20
REPORT.HS= REPORT 20
SILENCE. HS= SILENCE 20

65

TURN,HS = TURN\20

SPACE,HS = \20

SPEED,HS = SPEED\20

TAKE,HS = TAKE\20

ON.HS = ON 20

DECEPTIVE JTOUNTER_\IEASURES.I IS = DECM\20

WEAPONS_FREE,HS = WEAPONS FREE\20

WEAPONS JTIGHT,HS = WEAPONS TIGHT' 20

NUMBERS SET

USE,HS=USE\20
BACKSPACE.HS = \08
STATION.HS = STATION\20
ALL,HS = ALL 20

010 NUMBERS.CM
ONE,HS-l
THREE,HS = 3

FIVE.HS = 5
■ SEVEN.HS = 7
NINER,HS = 9
POINT.HS=.
SOUTH,HS = S\20
WEST,HS«W\20
ALPHA.HS = A
CHARLIE, HS = C
ECHO,HS = E
GOLF,HS = G
INDIA,HS=I
KILO.HS=K
MIKE.HS=M
OSCAR,HS = 0
QUEBEC,HS = Q
SIERRA,HS = S
UNIFORM, I IS = U
WHISKEY,HS = W
YANKEE.HS = Y
SPACE,HS = \20

AVIATION SET

TWO,HS-2

FOUR.HS-4

SIX.HS = 6
EIGHT,HS = 8
ZERO,HS = 0
NORTH,HS = N20
EAST,HS=E20
TACK,HS = -
BRAVO,HS=B
DELTA, I IS =D
FOXTROT,HS=F
HOTEL, I IS =11
JULLIET,HS = J
LIMA,HS=L
NOVEMBER,IIS = N
PAPA,HS=P
ROMEO,HS=R
TANGO,IIS = T
VICTOR. HS.= V
X-RAY.HS-X
ZULU,HS = Z
BACKSPACE,HS = \08

66

Oil AVIATION.CM

ALTITUDE,HS = ALTITUDE\20
BEARING, HS = BEARING 20
POSITION. HS= POSITION 20
BINGO.HS = BINGO\20
COURSE.HS = COURSEi20
FIRE.HS = FIRE'20
LAUNCH, HS= LAUNCH, 20
AEW,HS = AEW 20
ASW.HS = ASW(20
RECONN.HS = RECONN 20
RESCUE,HS= RESCUE 20
STRIKE_CAP.HS = STRCAP 20
SURCAP.HS= SURCAP 20
JAMMER,HS = JAMMER, 20
NONE.HS = NONE\20
SPEED, HS= SPEED \20
PROCEED,HS = PROCEED\20
STOP,HS = STOP\20
FOR,HS = FOR 20
CH46,HS = CH46\20
E3A,HS = E3A 20
EP3E.HS = EP3E 20
FA18.HS=FA1S,20
P3C.HS=P3C 20
LAMPS.HS=SH2F 20
SPACE,HS = \20

BARRIER,HS= BARRIER\20
FORCE.HS= FORCE' 20
TRACK,HS = TRACK\20
TO,HS = TO\20

COVER,HS = COVER' 20

AT,HS = AT\20

MISSION, HS= MISSION 20

AIRTANKER.HS = AIRTANKER 20

DECOY, HS= DECOY 20

RELAY.HS= RELAY 20

SEARCH. IIS = SEARCH 20

SURVEILANCE,HS = SURVEI LANCE 20

CAP.HS = CAP 20

STRIKE.HS = STRIKE 20

REFUEL, HS= REFUEL 20

TAKE,HS=TAKE\20

STATION, HS= STATION\20

A6E,HS = A6E\20

A7E.HS = A7E 20

E2C,HS = E2C\20

EA6B,HS = EA6B 20

F14A,HS=F14A 20

KA6D,HS=KA6D 20

S3A.HS = S3A 20

SH3H.HS = SH3H 20

BACKSPACE.HS = \03

SHIPS SET

012 SHIPS.NS = COMMANDS_TO_UNTTS.CT= 1,CM
KITTYHAWK,HS= FOR KITTY 20
FOX.HS=FOX 20
\VILSON.HS= FOR WILSO ,20
SPRUANCE.I1S= FOR SPRUA\20

67

KNOX,HS= FOR KNOX\20
WONSAN,HS= FOR WONSA\20
LOS_ANGLES,HS=FOR LOSAN\20
MISSAWA,HS= FOR MISAW\20
ADAK,HS= FOR ADAK\20
JFK,HS=FORJFK\20
R.K.TURNER,HS= FOR TURNR\20
MAC.HS=FOR MAC\20
FL"RER,HS = FOR FURER\20
IOWA,HS= FOR IOWA\20
LONGBEACH,HS=FOR LONGB\20
IOWA.HS=FOR IOWA 20
GAR,HS=FOR GAR 20
PETRO.HS= FOR PETRO.20
OMAHA,HS= FOR OMAHA\20
JOHN_ROGERS,HS=FOR ROGER\20
RATHBOURNE,HS=FOR RATHB\20
WICHITAU,HS= FOR WICHI\20
ALEKSIUV,HS = FOR ALEKS\20
VLADIVOSTOK'S - FOR VLAD\20
MCCORMICK,HS= FOR MCCOR\20
JOHN_HANCOCK,HS= FOR HANCK\20

LOAD SET {WEAPON SET)

013 LOAD, HS= LOAD 20,CM
HARPOOX,HS = HRPON\20
TLAM,HS = TLAM\20
ASROC,HS = ASROC\20
MARK46A,HS=MK46A\20
MARK57,HS=MK57\20
MARK83,HS=MK83\20
76MILLIMETER,HS=MM76\20
ROCKEYE.HS= RKEYE\20
SPARROWS = SPAR ,20
WALLEYE, US = WALLI 20

TASM,HS = TASM 20
APAM,HS = APAM 20
MARK46,IIS=MK46 20
MARK48,HS=MK48\20

MARK82,HS=MK82 20
MARKS4,HS=MK84 20
PHIONEX,HS = PIIENX\20
SHRIKE, IIS = SHRIK 20
SIDEWINDER,HS = S\YDR 20
SM2ER,HS=SM2ER 20

68

ONE,HS=l TW0,HS=2

THREE,HS = 3 F0UR,HS = 4

FIVE,HS = 5 SIX.HS = 6

SEVEN. HS = 7 EIGHTHS = 8

NINER,HS = 9 ZERO.HS = 0

PINGER.HS = SSQ47\20 DIFAR.HS = SSQ53\20

DICASS.HS = SSQ62 20 SPACE. HS = \ 20

BACKSPACE. HS= OS

STANDARD_EXTENDED_R.ANGE.HS = STDER 20
STANDARD MEDIUM RANGE. HS = STDMR 20

69

LIST OF REFERENCES

Barr, Avron, and Edward A. Feigenbaum, 1981: The Handbook of Artificial
Intelligence, Vol. 1, HeurisTech Press, 409 pp.

Druzhinin, V. V., and D. S. Kontorov, 1972: Foreword to the Russian Edition of

Concept, Algorithm, Decision (A Soviet View), by S. M. Shtemenko (General of
the Army. L.S.S.R.). Superintendent of Documents. U.S. Government Printing
Office, Catalog No. D30l.79:6, Stock No. 008-070-0034409.

Dupuv. Col. T. N.. USAj Ret., 1986: In Search Of An American Philosophv of

Command and Control. A Preliminary Draft, Class Notes OS3636, Summer
Quarter, 25 pp.

Foley, James D.. Victor L. Wallace, and Peggv Chan, 1984: The Human Factors of

Computer Graphics Interaction Techniques. IEEE, Computer Graphics and
Applications. November, 13-43.

Harris. C. L. Lane. P. Shaha. and J. Tombrclla. 1983: NWISS, Naval Warfare

Interactive Simulation System I. sers Manual. U arlab Handout. Naval
Postgraduate School. 28 November. 17 pp.

Kavaler. Robert A.. 1986: The Design and Evaluation of a Speech Recognition System
for Engineering Workstations. Ph.D. dissertation, University "of California,
Berkeley, 162 pp.

Kurzweil, Raymond, 1984: The Coming Age of Intelligent Machines or "What is 'AL
Anyway?". Keynote Address The Institute of Electrical and Electronics Engineers
{IEEE)' International Conference on Computer Design, 14 pp.

Local Command Center Network {LCCN) Statement of Work for Request for Proposal,
23 October 1978.

Miller, G. A., 1956: The magical number seven, plus or minus two: Some limits

on our capacity for processing information. Psychological Review, 63, 81-97.

Murveit. Hv and Donald Bell. 1986: Speech Entry to a Natural- Language- Accessed

Data Base. SRI Project 6096, Contract "N0O39-83-K-0442. \tenlo Park, CA:
SRI International, 75 pp.

Naval Ocean Svstems Center (N'OSCL 1985: Navy Exploratorv Development
Program FY 86 Block Plan. Combat Direction, N02C, F5 August. 1-37.

Orr, George E.. 1983: Combat Operation C~ T. Fundamentals and Interactions.
Maxwell Air Force Base, AL: Air University Press, 99 pp.

Pallett, David S.. 1985: Performance Assssment of Automatic Speech Recognizers.
Journal of Research of the National Bureau of Standards. 90. 5, 371-387.

Poggio. A., J.J. Garcia Luna Aceves, E.J. Craighill. D. Monm, L. Aquilar, D.

Worthington. and J. Hight, SRI International. 1985: CCWS: A Computer-
Based Multimedia Information System. IEEE, October, 92-103.

Poock, Gary K.r1980: Experiments With Voice Input For Command and Control:

Using Voice Input To Operate A Distributed Computer Network. Naval
Postgraduate School, Monterey, CA, 34 pp.

Poock, Garv K., 1981: A Longitudinal Study of Computer Voice Recognition

Performance and Vocabulary Size. Naval Postgraduate School. Monterev, CA,
32 pp.

Poock. Gary K., 1986a: A Longitudinal Study of Five Year Old Speech Reference
Patterns. Journal of the "American Voice Input: Output Society. 3, 13-18.

Poock. Gary K., 19S6b: Speech Recognition Research, Applications and International

70

Efforts. Invited Paper for the 1986 Human Factors Society, not yet published.

Ryan. Jr.. Thomas A., Joiner, Brian L., and Barbara F. Rvan. 1981: Mini tab Reference
Manual. University Park, PA, The Pennsylvania State University, 154 pp.

Salfer, D. L., 1985: Voice Automation of Ship Control. Master's Thesis, Naval
Postgraduate School. Monterey, CA, September, 59 pp.

Strategic Computing Program, 1985: Chapter 2. Integration, Transition,

and Performance evaluation of Speech Technology. Draft Copy. December.

Tanenbaum, A. S., 1981: Computer Networks. Prentice-Hall. Inc., 517 pp.

U.S., Joint Chiefs of Staff. Publication Number 1, 1984: Department of Defense

Dictionary of Military and Associated Terms. U.S. Government Pontine Office.
404 pp.

VTR 60x0 Series II. 1985: VOTAN Manufacturer's Technical Manual, Guide To
Procedures. VOTAN, Fremont, CA, 81 pp.

VTR 6050 II. 1985: User's Guide. VOTAN Manufacturer' s Technical Manual,
Reference Manual. VOTAN. Fremont. CA, 178 pp.

Wohl. Joseph G.. 1981: Force Management Decision Requirements for Air Force

Tactical Command and Control. IEEE Transactions on Systems, Man. and
Cybernetics. SMC-1 1, 9. 618-639.

71

INITIAL DISTRIBUTION LIST

No. Copies

1. Defense Technical Information Center 2
Cameron Station

Alexandria, VA 22304-6145

2. Librarv. Code 0142 2

Naval Postgraduate School
Monterey, CA 93943-5002

3. Defense Contract Audit Agency 1
Cameron Station

Alexandria, VA 22314

4. Garv K. Poock. Code 55Pk 30
Naval Postgraduate School

Monterey, CA 93943-5000

5. Mr. Kennv Avila I
Lockheed Aircraft Corp.

Dent. 1-332
P.O.Box 33
Ontario, CA 91761

6. Dr. Janet Baker 1
Drason Systems

Chapel Bridge Park
55 Chapel Street
Newton, MA 02158

7. Dr. Sarah Blackstone 1
ASLHA

10801 Rockville Pike
Rockville, MD 20852

8. Donald Bell 1
SRI International

333 Ravenwood
Menlo Park. CA 94025

9. Hv Murveit 1
SRI International

333 Ravenwood
Menlo Park, CA 94025

10. Bill Lee 1
LITTON Svstems

M S 2-01 '
5115 Calvert Rd.
College Park, MD 20740

11. Leon Lerman 1
LMSC - 0 86-60. BT53

P.O. Box 3504
Sunnyvale, CA 94088

12. B. Jay J. Maritn 1
Perceptronics

21111 Irwin St.

Woodland Hills. CA 91367

73

13. Paul A. Manoione 1
Speech Svstems Incorp.

1§356 Ox'nard St.
Tarzana, CA 91356

14. Dr. David Pallett 1
National Bureau of Standards

A216 Technology
BLDG. 625
Gaithersburg, MD 20899

15. Ralph Pettit 1
Cubic Defense Svstems

9333 Balboa Ave.
San Diego. CA 92123

16. LCDR. Michael A. LeFever, USX 1

579B Wilkes Lane
Monterey. CA 93940

17. CDR. J. Stewart. Code 55ST 2
Naval Postgraduate School

Monterey. CA 93943-5000

IS. C3 Academic Group. Code 74 2

Prof. M. K. Sovereign
Naval Posteraduate^School
Monterey, CA 93943-5000

19. Naval War College 1
Wargaming Department

Sims Hall.TIO?
Newport, RI 02S41-5010

20. MAJOR T. J. Brown, USAF 1
Code 39

Naval Postgraduate School
Monterey, CA 93943-5000

21. Headquarters Rome Air Development Center AFSC 2
Office of the Chief Scientist

Attn: Fl Diamond
Griffiss AFB, NY 13441

22. Commander, Code 421 1
Naval Ocean Svstems Center

San Diego, CA' 92152

23. USA CFXOM 1
AMSEL-SEI-F

Attn: Dr. Israel Mavk
Ft. Mammouth, NJ '07703

24. National Defense University 1
Fort Lesley J. McNair

Washington, DC 20319

25. Michael J. Zyda. Code 52 1
Naval Postgraduate School

Monterey. CA 93943-5000

26. Frederick C. Johnson 1
4141 Jutland Drive

San Diego, CA 92117

27. Steve Nunn 1
Commander, Naval Ocean Svstems Command

Code 441

San Diego, CA 92152

74

1 7898 i„

30H00T

V0^

>4 AUG 90

t\yeS^

thesis
c.l

a 0 2 6*

221061

LeeftVfch recognition * «
SP 7 and control wot*
rtTtfon Environment.

2 ffcB 99

T
T

35723

B026P

-

Thesis

L458

c.l

2213S1

LeFever

Speech recognition in a
command and control work-
station environment.